Langdock Guide

Langdock is an enterprise AI platform that bundles Chat, Agents, Workflows and Integrations under one roof. Model-agnostic and EU-hosted. One login, one bill, one compliance review. Whether you work with Claude, GPT, Gemini or Mistral.

1.1 Platform Idea

Five building blocks on one platform:

Model agnostic: When a new model launches, it is usually available in Langdock within days. You pick per message in the model selector at the top left. No vendor lock-in at the model layer.

1.2 Security & Compliance

Langdock is built for enterprise use, not consumer use. That shows in the compliance properties:

• EU hosting: Microsoft Azure, Frankfurt region. Data stays in the EU.
• GDPR-compliant: Contractually backed. Data Processing Agreement (DPA) available.
• No training use: Inputs and outputs are not used to train the models. Contractually guaranteed.
• ISO 27001 and SOC 2 Type II certified.
• SAML & SCIM for single sign-on and automatic user provisioning (Microsoft Entra, Google, Okta).
• Audit logs & usage exports for internal reporting and external audits.

1.3 Deployment Modes

1.4 Who Uses Langdock

As of 2026, 5,000+ companies use Langdock with 60,000+ monthly active users in total. Notable examples:

• Pharma & life sciences: Merck
• HR-tech: Personio
• Fintech: SumUp
• Media: Der Spiegel

Typical use cases range from knowledge work (research, writing, analysis) through customer support, content production, compliance reviews, to specialized agents for internal processes.

Productive in five minutes. This chapter only shows where things sit. Content and use cases come from chapter 4 onwards.

2.1 Login & Workspace Tour

Login via browser or mobile app. If your company has SSO, click »Sign in with SSO« and you land in the right workspace automatically.

After login you see three areas:

• Left sidebar: chat history, search, inbox, projects, library, agents, workflows. Since July 2026, Integrations, Skills and Prompts are grouped under a »Customize« entry, and personal settings have moved to the bottom of the sidebar.
• Center: the chat interface. This is where the interaction happens.
• Top header: model selector (left), workspace switcher and profile (right).

Inside the chat you find a »+« icon next to the input field. Behind it sit the advanced features like Deep Research, Document Editor, web search, company knowledge, and file upload.

2.2 Account Settings

Profile picture bottom left → Settings. The most important knobs for you personally:

• Language: German, English, French, Spanish or Italian for the interface. French, Spanish and Italian were added in July 2026, with the web app fully localised in each.
• Theme: light, dark, or system-driven.
• Notifications: email and in-app notifications for shared content, workflow results, usage limits.
• Custom instructions: who you are and how you want answers. Details in chapter 8.

2.3 Get Started Guide, Desktop & Mobile App

Get Started Guide: Langdock's own interactive onboarding. A checklist with tasks across chat, prompt library, agents, knowledge, integrations, workflows. With points and a workspace-wide leaderboard (admins can anonymize it). Three built-in helpers:

• Spotlight Tours: Interactive step-by-step UI guidance for many tasks. Shows you exactly where to click.
• Leaderboard: Shows your team's onboarding progress. A playful adoption lever.
• Custom Tasks: Admins can add their own onboarding tasks, e.g. »Read our AI usage policy«.

New users should run through it once. It does not replace this guide here, but it makes the UI tangible.

Mobile app: iOS and Android. Full chat, projects, actions, voice input. As of 2026 the app is on par with most desktop features.

Desktop app: Since 10 September 2026 Langdock is also available as a native app for macOS and Windows, downloadable at langdock.com/products/desktop (Mac ships as a .dmg you drag into Applications, Windows as a Setup.exe). It brings the same workspace as the browser: chat, agents, knowledge, workflows, integrations and workspace settings. On top of that it has its own icon in the dock or taskbar, multiple tabs for parallel tasks, keyboard shortcuts for common actions and several windows side by side. The first launch signs you in through the browser, after that browser and app can stay signed in at the same time. The app replaces the older Langdock app (PWA); minor differences around windows, menus and machine-level features are to be expected.

2.4 Keyboard Shortcuts

Three things to understand and the rest runs smoothly: which model when, how the usage limits work, how to save tokens without losing quality.

3.1 Model Families

Langdock gives you access to the major providers. Each has its strengths. You pick per message.

Which versions are available changes monthly. The model selector always shows the current line-up.

3.2 Three-Tier Model for Model Choice

Instead of guessing each time: three tiers, one rule of thumb.

Auto mode when you don't want to decide: Since May 2026 the model selector has an »Auto« entry. Langdock reads your first message, estimates complexity and picks a suitable model for the rest of the chat. Langdock decides which one and adjusts that over time, and you can see the choice in the model selector. It is built for everyday work: quick questions, drafting, summarising, rewriting and analysis. Good for routine chats and new users. For anything that really needs the top tier (Opus 5.5, the current top tier since late September 2026) or should run intentionally lean (Haiku 5.5, GPT-6 Luna), keep picking manually.

Use the reasoning toggle deliberately: A simplified reasoning toggle sits in the model selector (since May 2026). Activate it only when the task really needs reasoning. For strategy questions, architecture decisions, math, hard code problems. For standard text it produces longer answers and pulls more tokens without noticeably better quality.

For web research: GPT-5.5 is often cheaper and quality-comparable to Opus. Opus shines mainly on language and nuance. The extra spend is justified there.

New since late June 2026: Claude Sonnet 5. Anthropic's most agentic Sonnet yet, built for planning, tool use and autonomous execution. Quality approaches Opus 4.8 at noticeably lower cost per request, which makes Sonnet 5 the new default for daily work (EU-hosted). In workspaces where a previous Sonnet version was active, the model is enabled automatically.

New since July 2026: GPT-5.6. OpenAI's new flagship family supersedes GPT-5.5 and comes in three variants: Sol (flagship for coding and agentic search), Terra (balanced for daily work) and Luna (cost-efficient for simpler tasks). All three markedly improve knowledge work, reasoning and long-running agentic tasks over GPT-5.5 (EU-hosted). In workspaces where GPT-5.5 was active, they are enabled automatically. Prices dropped in August 2026: Luna costs roughly 80 % less, Terra roughly 20 % less than before. Langdock has since updated its recommendations for Auto mode and the cost fallback model, which makes Luna the cheapest sensible choice for simple tasks.

New since late July 2026: Claude Opus 5. Anthropic's new flagship supersedes Opus 4.8 as the top tier and steps up noticeably on complex reasoning, agentic coding, demanding knowledge work and long-running tasks (EU-hosted). In workspaces where a previous Opus version was active, the model is enabled automatically; the rollout can be controlled in workspace settings. Because Opus draws the most from the weekly limit, keep Opus 5 for genuinely high-quality tasks and let daily work stay on Sonnet 5.

New since mid-August 2026: Gemini 3.7 Flash. Google's strongest Flash model at the time, built for coding and agentic workflows, with gains in software engineering, web development and knowledge work. It is EU-hosted and has been enabled since 17 August 2026 in every workspace that already had a Gemini model active. Admins control access in workspace settings. Mainly interesting if you want to process large contexts cheaply: the Flash line remains the frugal tier, and its context window is considerably larger than that of the Claude and GPT models.

New since early September 2026: Claude Fable 5.1. Anthropic's most capable generally available family for coding and knowledge work, built for complex problem solving, long-running tasks and advanced research. EU-hosted and available since 1 September 2026. Where Fable 5 was already active, 5.1 is enabled automatically; admins control the rollout in workspace settings. Watch the budget: Fable is the most expensive tier in the catalogue. Fable 5.1 is listed at 8.94 euros per 1M input tokens and 44.70 euros per 1M output tokens (as of 15.10.2026), roughly twice as much as Opus 5 (4.47 and 22.35 euros) and about two and a half times as much as Opus 5.5 (3.58 and 17.88 euros). That makes Fable a fit for the few tasks where Opus 5.5 demonstrably falls short, not for daily work.

Also new since early September 2026: Gemini 3.8 Flash. Google's strongest Flash model so far, aimed at long-horizon software engineering, autonomous agents and complex enterprise workflows. Against 3.7 Flash it gains on coding, reasoning and knowledge work. EU-hosted and, since 2 September 2026, enabled automatically in every workspace that already had a Gemini model active, controllable via workspace settings. One note for cost planning: the price list on langdock.com/models lags a few days behind for brand-new models, so the changelog is authoritative.

New since early September 2026: GPT-6 Astra. OpenAI's most capable flagship so far, a step beyond GPT-5.6 Sol. Its strengths are complex reasoning, coding, computer use, browsing, scientific questions and demanding professional work. EU-hosted and available in all workspaces since 7 September 2026; where a previous GPT model was already active, Astra is enabled automatically. Day-to-day work stays on Sonnet 5 or GPT-5.6 Terra, Astra belongs to the genuinely hard tasks.

New since late September 2026: GPT-6 Sol, GPT-6 Luna and GPT-6.1 Sol. On 23 September 2026 OpenAI added two variants to the GPT-6 family: GPT-6 Sol is the balanced choice for complex coding and agentic workflows, GPT-6 Luna is built for focused, high-volume tasks. Both improve on professional work, coding and computer use and are available in every workspace. Since 30 September 2026 there is also GPT-6.1 Sol, which beats GPT-6 Sol at writing and debugging code, understanding complex documents and completing multi-step business workflows; where a GPT model was active before, it is enabled automatically. All three are EU-hosted. For cost planning: GPT-6 Luna is listed at 0.10 euros per 1M input tokens and 0.49 euros per 1M output tokens, cheaper still than GPT-5.6 Luna (0.20 and 1.18 euros). GPT-6 Sol and GPT-6.1 Sol both sit at 1.97 and 9.83 euros, roughly on par with Sonnet 5.5, while GPT-6 Astra at 8.94 and 44.70 euros is on par with Fable (all prices as of 15.10.2026). Since October 2026 the fair usage docs name GPT-6 Luna as the fallback model when you exceed your limits (Langdock Cloud without BYOK, section 3.5).

New since late September 2026: Claude Opus 5.5 and Claude Sonnet 5.5. Opus 5.5 (since 23 September) improves on Opus 5 in agentic coding, knowledge work and communication while also being cheaper: the price list shows 3.58 euros per 1M input tokens and 17.88 euros per 1M output tokens, against 4.47 and 22.35 euros for Opus 5 (as of 15.10.2026). Sonnet 5.5 (since 28 September) is stronger at everyday coding and knowledge work and produces more polished documents, slides and spreadsheets. It is enabled automatically wherever a Sonnet model was active before. It is now on the price list at langdock.com/models, at the same price as Sonnet 5 (1.79 / 8.94 euros per 1M input and output tokens, as of 15.10.2026), so switching costs nothing extra. Both are EU-hosted. This shifts the recommendation in 3.2: Sonnet 5.5 for daily work, Opus 5.5 for high-quality tasks.

New since October 2026: Claude Haiku 5.5 and Gemini 4 Argon. Haiku 5.5 (since 8 October) is Anthropic's small, fast model for high-volume work such as summaries or tasks an agent hands off to subagents. Where other Anthropic models are already active, it is enabled automatically. The price list shows 0.10 / 0.49 euros per 1M input and output tokens, about a ninth of Haiku 4.5 (0.89 / 4.47 euros) and level with GPT-6 Luna. Haiku 5.5 therefore replaces Haiku 4.5 as the recommendation for the routine tier in 3.2. Gemini 4 Argon (since 13 October) brings stronger reasoning and writing for demanding professional tasks, from debugging and refactoring code to understanding detailed documents and running multi-step business workflows. Where an earlier Gemini model was active, it is enabled automatically. At 1.97 / 9.83 euros it sits on the GPT-6 Sol tier and costs more than twice as much as Gemini 3.8 Flash (0.74 / 3.69 euros), so Flash remains the leaner choice for quick routine work. Both are EU-hosted and available in every workspace; admins control access in workspace settings. Prices as of 15.10.2026.

3.3 Special Variants: Vision, Reasoning, Image

• Vision (image understanding): Sonnet, Opus, GPT-5.x, Gemini Pro can read images. Attach them like a file.
• Reasoning models: o3, o4-mini (OpenAI) and the »Thinking« variants of GPT-5/Claude/Gemini think first, then answer. For multi-step logic, math, hard code problems.
• Image generation: GPT-Image 2.5 (Flare and Sunburst variants), Imagen, Flux. Via the »+« menu in chat or as a tool inside a workflow.

3.4 Session & Weekly Window

Since April 2026 Langdock works with two parallel usage windows, comparable to ChatGPT Plus or Claude Pro:

Both run in parallel. Whoever exceeds one triggers the fallback mechanic.

Consumption is weighted by provider cost: Opus pulls more than Sonnet, Sonnet more than Haiku or GPT-Mini. One Opus call equals roughly two to three Sonnet calls.

3.5 What Happens When You Exceed

• New messages are automatically routed to GPT-6 Luna until the window resets.
• In-flight responses finish before the system switches.
• You don't get blocked. The workflow continues, just on a cheaper model.

3.6 Spam Protection

Independent of cost-based limits there is a hard cap of 250 messages per 3 hours. It targets automated abuse, not normal use. Manual chatting never gets close.

3.7 See Your Own Usage

Two places show usage:

• In the model selector: Since May 2026 usage bars sit right next to each model name. As you switch between models you immediately see how full the session and weekly windows are. No detour through Settings.
• Settings → Account → Usage: Shows two progress bars (session and week) plus the time until each reset.

Anyone often near the limit should walk through 3.2 (three-tier model) and 3.9 (chat hygiene) before booking more capacity.

3.8 Limit Cascade: More Capacity

For sustained more capacity, six stages are available, ordered from »costs nothing« to »costs extra«:

Admin configuration details in chapter 20.

3.9 Chat Hygiene: Saving Tokens Without Quality Loss

Four levers every user has themselves. More effect than any plan upgrade.

• Start a new chat when the topic shifts or after 10 to 15 turns at the latest. The full history goes into the model as input on every turn. The longer the chat, the more input tokens. Prompt caching softens the effect but does not remove it.
• Keep custom instructions lean. They get sent as the system prompt on every turn. Long instructions cost tokens and often hurt response quality (instruction overload). Stay focused on the essentials.
• Attach files deliberately. Instead of uploading a 50 MB PDF, cut out the relevant sections. Or: drop them into a folder and let semantic search pull pieces on demand instead of pushing the whole document into context (chapter 12).
• Cap output length deliberately. »Max 5 bullets«, »in 100 words«, »no preamble« saves tokens and usually makes the response better.

The daily tool. Most tasks land here, from drafting an email through analyzing a spreadsheet to sketching a diagram. All advanced features sit behind the »+« icon next to the input field.

4.1 Sending, Editing, Branching

• Send: Enter. Shift + Enter for line break without sending.
• Edit your prompt: Pencil icon on hover over your prompt. The response regenerates from there.
• Regenerate response: Circle-arrow button under the answer. Generates alternatives without changing the prompt.
• Copy response: Clipboard button. Stays visible while scrolling. Citations stay clickable as hyperlinks to the source when copied.
• Streaming reconnect: If you navigate away during a response and return, streaming continues. No loss. Since August 2026 long-running chats also hold the connection while the model is thinking or a tool is working and nothing comes back in between.
• Current time / timezone: Models have access to the current date and timezone via a built-in tool. You no longer need to pass »today« in the prompt.

4.2 File Attachments

Via the »+« icon or drag-and-drop. Supported: PDF, DOCX, Excel/CSV, images, code files, TXT, Markdown and more.

• Limit: 20 files per chat. For sustained needs, folders are the right path (chapter 12).
• Large PDFs: Better cut out the relevant sections than uploading a complete 50 MB document. Saves tokens and makes answers more precise.
• Excel/CSV: Combine with Data Analysis (chapter 11) for evaluations with real code execution.
• Images: Pick a model with vision (Sonnet, Opus, GPT-5, Gemini Pro). More in 4.5.
• Video: Since June 2026 video files can be uploaded up to 100 MB (previously 20 MB).
• Audio & video transcription: Since August 2026 Langdock turns uploaded audio and video files into transcripts with speaker labels. Contributions are attributed to the people who made them, which is what makes meeting recordings and interviews genuinely usable.
• Calendar & e-books: Calendar invites (.ics) and e-books (.epub) can be uploaded directly since August 2026, with no conversion step in between.
• Save directly: Files from chat can be saved straight to your Google Drive or OneDrive in a click, no local download detour.

4.3 Web Search

»+« menu → activate Web Search, or simply ask for it in chat. The model searches the web live and returns answers with source citations.

Admins can block domains workspace-wide (e.g. competitor sites). For deeper research see chapter 10 Deep Research.

4.4 Image Generation

»+« menu → Image Generation. Available models: GPT-Image 2.5, Imagen, Flux. Each has its own character.

• GPT-Image 2.5: OpenAI's current image generation since 11 September 2026, EU-hosted and enabled in all workspaces. Two variants: Flare as the default with better quality, better image editing and more speed, Sunburst as the premium option for maximum precision. Where GPT-Image 2 was active before, Flare is on automatically. Strong on text inside images and brand-consistent outputs.
• Imagen: Photo-realistic, good anatomy and composition.
• Flux: Fast, good for iteration and stylistic variants.

Iterate via re-prompts: »same scene but from above« or »warmer lighting«.

4.5 Image Analysis (Vision)

Attach an image as a file, pick a model with vision. Use cases:

• Have screenshots explained (»What is happening here?«)
• Evaluate diagrams (»What trends sit in the data?«)
• Extract text from images (receipts, whiteboards)
• Design reviews (UI, layouts, brand consistency)

4.6 Mermaid Diagrams

If you ask models to render something »as a Mermaid diagram«, they generate Mermaid code. Langdock renders it inline in the chat as a diagram. Handy for flowcharts, sequence diagrams, ER diagrams.

Render the onboarding process for new hires as a Mermaid flowchart.
Steps: contract review, IT setup, welcome meeting, first task.

Math formulas: Since June 2026 Langdock renders mathematical formulas (LaTeX notation) cleanly inline in the chat, across all models. Fractions, sums, matrices, and Greek symbols appear as typeset formulas instead of raw text.

4.7 Document Search

»+« menu → Document Search. Searches semantically across the attached or linked files. Different from a direct »answer from file« query, Document Search returns the passages plus a synthesized answer. Good when you want to see the sources.

Page previews for Word and PowerPoint: Since May 2026, Langdock renders visual page previews for DOCX, PPTX, DOC and PPT. Shorter documents show every page automatically, longer ones surface beginning, middle and end. Need more (e.g. for tables or diagrams)? Request them explicitly. Citations stay stable during streaming and remain clickable hyperlinks after you copy them.

4.10 Full-text Chat Search

Since May 2026, chat search no longer scans titles only, but also the body of messages. You can find chats by concrete keywords from the conversation. The sidebar also remembers its open/closed state across sessions.

4.8 @-Mentions: Agents, Workflows, Skills in Chat

In chat, a @ is enough to bring other building blocks in:

• Agents: Type @, the picker shows available agents. The selected agent answers with its own configuration (knowledge, tools, instructions).
• Workflows: Same trigger. The workflow runs and its result lands back in chat.
• Skills: Activate automatically when the context matches. No @ needed.

4.9 Company Knowledge

Since January 2026: »+« menu → Company Knowledge. One search across all connected data sources. Eleven integrations are covered today: Google Drive, SharePoint, OneDrive, Confluence, Gmail, Outlook, Slack, Microsoft Teams, Linear, Jira and Google Calendar. It searches documents, spreadsheets, emails, chat messages, tickets and meetings.

• Cross-source: No separate search per system. Langdock breaks the question apart, searches the selected sources in parallel, ranks by relevance and writes one answer from the results.
• Pick sources per search: Activating it opens a selector with all connected tools. You toggle them individually, at least one has to stay on. Multiple accounts per service are supported.
• Permissions from the source system: If you don't have access to file X, you don't see it here either.
• Traceability: A timeline panel shows live which system is being searched, plus tabs with the search path and direct links to the source documents.
• Setup: An admin enables Company Knowledge in the workspace settings, then each user connects their own systems (chapter 14).

4.11 Keep an Eye on the Context Window

Since June 2026 an indicator in the input bar shows how full the current chat's context window already is, e.g. »13.6k / 200.0k (7%)«. Clicking it opens a breakdown by category: system tools, system prompt, messages, tools, attachments and images, skills, memory, autocompact summary, and free space. Each category shows token count and share.

Handy when answers in long chats start forgetting the beginning of the conversation: you see exactly what is eating the space (often old attachments) and can clean up selectively or start a fresh chat instead of writing on blindly.

4.12 Scheduled Tasks

Since June 2026 you can have a prompt run automatically on a recurring basis, daily, weekly, or on a custom cadence. A scheduled task fires its prompt at the set time, pulls fresh data through connected agents and integrations where needed, and stores the result as a run you can review later.

• Create: Write the prompt, pick the cadence (e.g. every Monday at 8am). Optionally attach an agent or a connection so the task fetches current data.
• Cadence: Manual (on demand), daily, weekdays (Mon to Fri), weekly on a fixed day, on selected weekdays, or monthly. In short months a monthly run falls back to the last day automatically.
• Create from an ongoing chat: Since August 2026 you no longer have to switch to the Scheduled section. Ask in chat to have something run regularly, Langdock proposes name, instructions and cadence, and you approve. Tasks created this way use your browser's timezone and run in Auto mode where available.
• Limit: Up to 10 scheduled tasks per user per workspace. Runs may start up to 15 minutes late because Langdock spreads executions across the window.
• Preset gallery: Since July 2026 there are ready-made presets for scheduled agents, with name, instructions and cadence prefilled. You only adjust them instead of starting from scratch.
• Own section: A dedicated »Scheduled« section in the sidebar lets you manage and monitor all tasks in one place.
• Approving actions: If the task needs an action that requires confirmation, you approve it for the single run or permanently via »allow for future runs«.
• History: Every past run stays available, so you can read back the results of earlier executions and trigger a run manually at any time. Skipped runs do not create a chat.
• Pause: Tasks can be paused and resumed later without recreating them.
• Templates stay put: Since mid-September 2026 the templates are permanently visible in the task list instead of only appearing when you create a task. New runs can also be cleared in one go via »mark all as read«.
• Controllable via API: Also since mid-September 2026, scheduled tasks are available through the API: create, list, update, pause, resume, run now and delete. Useful when the tasks are maintained from another system (see chapter 19.2).

Three levers cover 90 percent of tasks. You only need depth for special cases.

5.1 Clear Instructions, Role, Structure

Tell the model precisely what you want, in which role, with what knowledge. A proven structure:

• Persona: Who should the model be? (»You are a senior marketing manager.«)
• Task: What should happen? (»Write a LinkedIn ad.«)
• Context: What knowledge is relevant? (»Audience: SaaS founders.«)
• Format: What should the result look like? (»Three variants, max 100 words each.«)
• Examples (optional): One or two sample outputs.

Four out of five usually does the job. Examples deliver the biggest quality jump because the model sees exactly what you want.

5.2 Control the Output Format

Without an explicit ask, you get whatever the machine considers appropriate. Often too long, too generic, with preamble.

• Form: »in bullets«, »as a table«, »as JSON with fields X/Y/Z«, »as Markdown«.
• Length: »in 100 words«, »max 5 bullets«, »one sentence per item«.
• Style: »no preamble«, »no disclaimer«, »in second person«.

5.3 Chain Prompts & Tips

Break complex tasks into steps instead of cramming them into a mega-prompt:

1. First: »Collect 10 arguments for X.«
2. Then: »Rate each argument on a scale of 1 to 5.«
3. Then: »Write the pitch with the three strongest.«

Two unusual but effective tricks:

• Avoid »don't« instructions. Models handle negations worse than positive instructions. Instead of »Don't write in marketing speak« better »Write directly and without sales floskel.«
• Ask for direct quotes. When you want facts from a document, ask explicitly »Quote the relevant passage verbatim.« Reduces hallucinations.

Save recurring prompts instead of rewriting them every time. Low effort, high payoff, especially for teams.

6.1 Saving, Variables, Folders

• Save: When hovering over your prompt a »+« icon appears → save to library. Title and tags.
• Variables using {{name}} notation make templates flexible: »Write a recap for the meeting with {{customer}} on {{date}}.«
• Folders for organization. A useful structure: by function (marketing, sales, support) or by task (recap, outreach, research).

6.2 Sharing With Team, Using in Chat & Agents

• Sharing: With individual users, groups, or the whole workspace. Workspace-wide prompts show up for everyone.
• In any chat: Slash command / or picker top right. Variables get prompted in a dialog before the prompt is sent.
• In agents: Also usable as default prompt for an agent. Everyone gets the same starting point.

Projects are containers for related work. When a single chat is too small and an agent too big, the project is the right tool.

7.1 Project Concept

A project bundles three things under one roof:

• Multiple chats on the same topic (e.g. »Q3 reporting« with research chat, draft chat, review chat).
• Shared files available in every chat of the project (briefing, templates, guidelines).
• Own custom instructions active in every chat of the project.

Examples from practice:

• Q3 reporting: Excel data + briefing + custom instructions »results are for the board meeting«.
• Pitch for customer X: recent pitches as templates + branding guidelines + custom instructions »write in second person, not too marketing-heavy«.
• Onboarding a new vendor: contract + standard question catalog + custom instructions »always check against our T&Cs«.

7.2 Creating, Attaching Files Permanently

Left sidebar → Projects → new project. Name, optional description. Then upload files. They are available in every chat of the project from now on.

Files can be swapped later without losing the running chats. That is the main advantage over direct attachments per chat.

7.3 Custom Instructions at Project Level

In project settings you place instructions that work in addition to your personal custom instructions (chapter 8). Examples:

• »I create outputs for the board meeting here. Keep the tone factual.«
• »Always answer in English, all other projects in German.«
• »Back every recommendation with a source from the attached documents.«

7.4 Project Sharing

Since December 2025 projects can be shared with users, groups, or the whole workspace. Anyone with access sees all chats within the project and works on the same file basis.

• Reader: Read + start new chats, but cannot change project settings.
• Editor: Full rights including swapping files and adjusting custom instructions.

7.5 Project vs. Agent: When What

Several mechanisms keep the model context-aware. It knows who you are, what the company does, what project is running. Important: don't mix them up. Each layer has its own place.

8.1 Memory

The model automatically remembers relevant facts from conversations when you signal it (»note: ...«) or when the system considers the information important enough.

• View: Settings → Account → Memory. List of all entries.
• Edit / delete: Per entry, anytime. If an old entry is wrong, you can correct it manually.
• Disable: Switch off entirely or per chat (»forget this right after«).

8.2 Three Levels of Custom Instructions

All three are sent as system prompt with every chat. The effect: every employee gets »on brand« outputs immediately, without explaining the company every time.

8.3 Memory vs. Custom Instructions vs. Project Instructions

Since May 2026 the former Canvas mode is called Document Editor. Rebuilt from scratch with richer formatting, version history, cloud export and Library integration. Existing Canvas content stays accessible inside its original chats, and prompts mentioning »Canvas« still open the editor.

9.1 What the Document Editor Is

A side panel next to the chat where you edit text or code directly. Instead of routing every change through the dialogue, you write inside the document and let the AI assist in a targeted way.

• Three ways to start: via prompt (»Draft a proposal about X«), via the »Create document« button with Document/Word/PDF choices, or through the Tools menu.
• For code: Prompts like »Build me a budget tracker« or »Create a bar chart from this data« automatically open code mode.
• Full screen: An expand icon hides the chat panel when you want to focus only on the document.

9.2 Formatting & Slash Menu

Press / to open a menu with three groups:

• Format: Headings (H1 to H4), text variants.
• Lists: Bullet and numbered lists, blockquotes, code blocks.
• Insert: Tables, auto-generated table of contents, dividers.

A persistent toolbar replicates the same options if you don't want to invoke the slash menu every time.

9.3 AI in the Editor: Targeted or Sweeping

• Targeted edit: Highlight text and pick »Edit with AI«. Only that passage changes.
• Sweeping change: Without selection, send an instruction in the chat (»cut to half« or »add a conclusion«). The AI regenerates the whole document accordingly.

9.4 Code Mode with Live Preview

In code mode you can switch between two views:

• Code: Read and edit the source file.
• Preview: Live rendering to test the output right away.

Handy for interactive mini-apps, diagrams or UI snippets you want to see immediately.

Fix with AI on runtime errors: When the preview throws a runtime error, a »Fix with AI« button appears right next to the error. One click and the model parses the message and patches the code. Saves manual stack-trace reading, especially for snippets in languages or frameworks you don't live in.

9.5 Export & Cloud Storage

• Text documents: Export as Markdown, Word (DOCX) or PDF.
• Code files: Download in native formats (.html, .jsx, .tsx).
• Direct save: One click to push the file to SharePoint, OneDrive or Google Drive. Skips the download-and-upload detour.

9.6 Version History & Library

• Versioning: Every change (from you or the AI) gets a timestamp and an attribution. The version panel shows diffs and lets you roll back to any previous version without losing manual edits.
• Library integration: All created documents and code files land in the Library automatically (chapter 13). You find them across chats and can keep working on them from any conversation.

9.7 When Document Editor, When Regular Chat

Deep Research is the mode when a simple web search isn't enough and you want multi-stage research.

10.1 What Deep Research Is

»+« menu → Deep Research. The model starts a multi-stage research run that can take a few minutes up to an hour. It collects sources, compares, synthesizes. The output is a structured report with citations.

10.2 When to Use, When Not

10.3 Limits & Models

• Own usage limits: Standard workspaces get 15 researches per user in a rolling 30-day window. BYOK workspaces set their own limit or remove it entirely. On top of that, Deep Research counts towards the regular session and weekly windows (chapter 3).
• Models: Three roles work together behind the scenes: a backbone model runs the search loop and polishes the final report, a reasoning model plans the research and evaluates results, and a fast reasoning model summarises the individual research tasks. Standard workspaces get Langdock's own configuration; BYOK workspaces supply the reasoning and fast reasoning models themselves.
• Runtime: Typically 5 to 30 minutes. You can keep working in the workspace meanwhile and get a notification when the report is done.
• Output: A structured report with citations and source links, plus a log of every search performed. The report can be downloaded as a PDF.

Data Analysis lets the model run real code in a sandboxed environment against the data you attach. Different from »normal« chat output the results are reproducible and verifiable.

11.1 Sandbox Execution for Excel, CSV, Code

»+« menu → Data Analysis. Attach an Excel or CSV file, formulate the task. The model writes Python code, executes it in a Jupyter sandbox, returns results plus the code. Pre-installed libraries include pandas, numpy, matplotlib, python-docx, and reportlab.

• Pivot, aggregation, joins, filters are standard.
• Data stays in the Langdock context. No transfer to third parties, no code runs on your machine.
• Multiple files can be combined (joins on shared columns).
• Sandbox limits: 60-second timeout per code run, no internet access from the sandbox, data is wiped after 15 minutes of inactivity. Session state is preserved within a single conversation.

11.2 Charts & Output Formats

• Charts: line, bar, heatmap, scatter, boxplot. Rendered inline in chat.
• Structured extracts as Markdown tables.
• Excel export of the result table via the download button.

11.3 Best Practices

• Make column headers clear. If columns are called »col_1«, »col_2«, the model has to guess what they mean.
• For large files: have it ask first what to check. Instead of »analyze the file« better »what evaluation do you suggest?« Then proceed deliberately.
• For multiple files: spell out the relationship. »File A is the master with all customers, file B has revenue per month. Join on customer_id.«
• Always read the code. When the model does something wrong, you see it in the code, not in the prose.

Folders and knowledge bases are the two ways to make knowledge and files accessible to AI answers without pushing whole documents into every chat context. Folders are the shared workspace, knowledge bases the searchable reference (see 12.7).

12.1 Folder Concept: the shared workspace

Since June 2026 folders are active workspaces, no longer plain search collections. You create a folder once, fill it with the relevant files, and work with it in any chat. The model reads the full contents, compares and updates files, and saves newly generated files straight back into the folder (as long as you have editor or owner rights). Typical tasks: summarizing several documents into a meeting briefing, analyzing spreadsheets, building a presentation from source material, or updating an existing document collaboratively.

A folder holds up to 100 files and 250 MB. For large reference collections where the model should only find the most relevant snippet semantically (up to 1,000 files), use a knowledge base instead (see 12.7).

12.2 Creating, Uploading Files

• Left sidebar → Library → Folders → new folder.
• Name and optional description (helps with search later).
• Files via drag-and-drop or browser upload. Multiple at once.
• Uploaded files get indexed. That takes seconds (TXT) up to minutes (large PDFs).

Permissions at file level: anyone with folder access can search all files inside. For cross-source mechanics with source-specific permissions see chapter 4.9 Company Knowledge.

12.3 Sharing and Roles

Folders can be shared with individuals, groups, or the entire workspace (often used for company knowledge). For each person you assign one of three roles. Key point: only editors and owners may save model-generated files back into the folder.

Workspace admins have access to all folders across the board.

12.4 Folders in Chat, Projects and Agents

• In chat: the »Work in a folder« button (or »+« menu → folder) and pick a folder. The model then sees all files in the folder and can save generated files back (with editor/owner rights). Since September 2026 this also works after the fact: an already running chat can be moved into a folder through the same option, so you no longer have to start over.
• In a project: attach a folder to a project when you own both. Project permissions then govern folder access, and the team uses the folder contents together as context in project chats.
• In an agent: assign one or more folders during configuration. The agent reads the files permanently but cannot modify or save anything back; management stays with the folder owner.

12.5 Folder vs. Direct Attachment vs. File Templates

12.6 File Types, Limits, FAQ

• Supported types: PDF, DOCX, XLSX, PPTX, TXT, MD, HTML, CSV and more.
• Table limits: XLSX/CSV are indexed semantically. For real data analysis use chapter 11 instead.
• Code repositories as a whole are not supported. Single-file code is fine.
• Images in a folder: Since July 2026 the model can read and analyze images in a folder visually, not just their file names. Screenshots, diagrams or photo templates can be evaluated straight from the working directory.
• Attachments per message: on the web you attach up to 50 files to one message, on mobile up to 5. Attachments share the context window with your prompt and chat history, so only send what you actually need.
• Agent knowledge and the 20-document threshold: an agent takes up to 50 files in its Files row. Up to 20 documents all sit in the model's context as a preview. Above that, Langdock switches mode: text documents stay reachable through file lists and file tools, spreadsheets stay directly readable. If you need more, attach a knowledge base instead of individual files (see 12.7).
• Folder capacity: a folder holds up to 100 files and 250 MB. Library uploads were doubled in mid-September 2026 from 50 to 100 files per upload. On top of that there is a 2 GB upload quota per 24 hours, and individual deployments can be configured higher. For larger holdings the knowledge base is the right place (see 12.7).
• Refresh: When you replace a file, it is reindexed. With Folder Sync (chapter 14) this happens automatically.

12.7 Knowledge Bases: the searchable reference

A knowledge base is a searchable document collection for reference material. Instead of loading a whole document into the model context, vector search passes only the most relevant section per query. That lets you connect significantly more material without blowing the token budget. In chat you call it up with »@« and select the knowledge base, then ask your question. Agents, workflows, and the Knowledge Folder API can access it too.

• Limits: up to 1,000 files, up to 8 million characters per file.
• Supported types: PDF, DOCX, TXT, Markdown and (text-based) presentations. Not supported: tabular files, images, audio.
• Which when: a folder when you want to open, compare, update, or save files back in chat. A knowledge base when you need to locate specific information in a large collection.
• Your own vector database instead of a knowledge base: if you already run your own retrieval infrastructure, connect it as an integration: Qdrant, Pinecone, Azure AI Search, Milvus and Vertex AI Vector Search are supported. Once connected, the database is available as an action in agents and via »@« in chat, and every request triggers an embedding search. Knowledge bases work with fixed defaults internally: chunks of roughly 2,000 characters, 1,536 dimensions, the top 50 matches. Your own database pays off once you are dealing with several thousand documents, or when you want to control chunking, metadata and the embedding model yourself. The price: operations, maintenance and access control are then on you. There is no automatic synchronisation from integrations into a knowledge base either way.
• Convert later: Since July 2026 a knowledge base can be converted into a regular folder if you decide you want to work on the files actively. You need editor or owner access: open the knowledge base in the Library, pick »Convert to folder« from the actions menu, confirm. Files and sharing permissions move across, and the original knowledge base is deleted. Conversion is blocked while the knowledge base is still connected to APIs, agents, chats or automations, so remove those connections first.

12.8 Best Practices

• Use descriptive filenames. Vector search uses contents, but the model shows file names in answers. »contract-2024-finalfinal-v3.pdf« doesn't help.
• Clear versioning. If old versions remain in the folder, the model answers with stale content. Better delete cleanly than keep them in parallel.
• Drop a context note in the folder. A short README.md with »what is here, who is the audience« makes answers better.
• Split by function. Three specialized folders (contracts, sales pitches, onboarding) work better than one folder with everything.

Since the Library launch in April 2026, Langdock bundles all file-related functionality in one area: folders, file templates, recent files.

13.1 Library Concept

The Library is the central store for everything you want to reuse:

• Folders: shared workspaces for actively working on files (chapter 12).
• Knowledge bases: searchable collections for reference knowledge (chapter 12.7).
• File templates: reusable templates for outputs.
• Recent files: what you or your team used recently.

Reachable via the left sidebar → Library.

Since July 2026 the Library can be searched and filtered directly, and several files can be selected at once (batch selection), for example to move or delete them together.

13.2 File Templates

File templates are reusable schemas that serve as scaffolding for new outputs. Examples:

• Briefing template: with fields for client, goal, target audience, deliverables, timing.
• Sales pitch template: slides or sections in a fixed sequence.
• Recap template: structured note after a meeting (attendees, decisions, to-dos).

Templates can be shared team-wide. When used in a chat or agent, the model produces output directly in the template format. Since May 2026 file templates can also be selected directly inside agents so every agent run lands in the right shape. Template uploads became more robust at the same time and file types that previously failed are now accepted reliably. PPTX generation also supports React icons in slides since May 2026, which visibly lifts the polish of generated decks.

13.3 Folders Inside the Library

Folders and knowledge bases now sit structurally under »Library«. Since June 2026 folders are active workspaces (edit files together and save them back), while knowledge bases remain the searchable reference. Details in chapter 12. The path change (from »Knowledge« to »Library«) makes it easier for new users to find.

13.4 Recent Files

List of recently used files, your own plus shared ones. Useful when you can't remember which chat or folder a document lived in. Filters by type and time range. Since June 2026 the Library remembers your sorting and filtering preferences automatically, so you don't have to set them again on every visit. Since July 2026 you can additionally filter folders and knowledge bases by name, and select all files at once in the Library views.

13.5 Creating Files Directly From the Library

In the Library section »Create New File« you start a document, spreadsheet, or presentation right away. Langdock opens a new chat in the matching mode, and if templates are enabled, recently used templates show up with a link to the full gallery.

Every file you create or upload surfaces with an in-app preview before it moves into a chat or folder. Saving to Google Drive or OneDrive happens with one click directly from the chat. Since July 2026 a file can be opened fullscreen straight from the preview dialog, which matters for presentations and multi-page PDFs, and since August 2026 zoom and a mobile-optimised layout come on top. Attachments and images in chat are directly clickable and easier to remove, and document previews refresh automatically once the AI has edited the document.

Integrations connect Langdock with the systems where work happens. SharePoint, Drive, Confluence, Salesforce, ServiceNow, Slack and dozens more. Integrations are the bridge between AI and the actual data and tool stack.

14.1 Integration Directory

Since mid-September 2026: plugins. Langdock groups integrations, skills and tools under the shared term plugins. You add them straight from the chat input instead of switching to settings first, which gives a conversation context and actions in one step. The sidebar lets you browse the catalogue by category and sort it by what is used most across the workspace, and every integration ships prompt starters as an entry point. For skills the previously separate install step is gone, they now have a single state: on or off (see chapter 15.3).

Left sidebar → Integrations, or directly from the »+« menu in chat. The directory shows:

• Native integrations: Pre-built, with tested actions and clean OAuth, now more than 100 of them. Beyond the classics (Microsoft 365, Google Workspace, Slack, Jira, Confluence, Salesforce, HubSpot) also project and task management (Asana, Linear, monday.com, Wrike, Microsoft Planner, Google Tasks, Microsoft To Do), notes and boards (Notion, Airtable, Miro, OneNote), calendars and meetings (Google Calendar, Outlook Calendar, Calendly, Luma, Google Meet, Zoom), development (GitHub, Bitbucket), service and support (ServiceNow, Zendesk, Pylon), data sources such as Snowflake, BigQuery, Databricks, Power BI, Tableau, Looker, Metabase, AWS Kendra and Azure AI Search, vector databases (Pinecone, Qdrant, Milvus, Vertex AI Vector Search) and specialist services such as DeepL, ElevenLabs, Statista, Stripe, Personio, Ashby, Shopify Admin, Microsoft Fabric and UiPath.
• MCP servers: Extensions connected via Model Context Protocol (chapter 18).
• Custom integrations: Self-built connections for internal systems (chapter 18).

The directory now covers more than 100 native integrations and keeps growing. Added in July 2026, among others: Pipedrive with 46 actions and 6 triggers across Deals, Persons, Organizations, Activities, Leads, Notes, Pipelines and Stages. Confluence (folders now sync into the Library), OneDrive (search also finds files in group drives and followed SharePoint sites) and Outlook Calendar (all-day events plus Exchange room booking) were extended as well.

August 2026 added Microsoft Viva Engage with 15 actions: browse communities, read and reply to messages, create announcements and polls, give praise, manage members. Existing integrations grew too: Confluence attaches files directly when creating and updating pages and can list and download attachments, ServiceNow returns the connected account's profile via »Get Current User«, Databricks gained seven actions around SQL queries, warehouses, notebooks and dashboards, Tableau can render views and workbooks as images, and ElevenLabs transcribes audio via »Speech to Text« and speaks text with the more expressive eleven_v3 model.

Also new in August 2026: Microsoft Fabric and Shopify Admin. Fabric connects business data through APIs for GraphQL. Langdock finds the accessible APIs, inspects the schema and runs focused queries; write requests go through a separate mutation action that has to be confirmed before it executes. Supported sources include Fabric Data Warehouse, Fabric SQL Database, Lakehouse and mirrored databases via their SQL analytics endpoints, plus directly connected Azure SQL Databases. Shopify Admin is connected with your own Shopify app: Client ID and Client Secret come from the Shopify Dev Dashboard and are stored as a custom OAuth client in the workspace admin settings. Worth planning for: the connection uses application-level permissions, so the app's Admin API scopes apply, not the Shopify role of the person who installed it.

Also in the directory since August 2026: UiPath and UiPath Automation Hub. UiPath lets you list folders and processes after an OAuth login and trigger jobs, so Langdock sets an existing RPA landscape in motion instead of rebuilding it. Automation Hub covers the reading side: search automation ideas, inspect flows, pull up cost-benefit analyses. Smaller additions landed for Snowflake (optional “Export as CSV” toggle), Jira (polling triggers return up to 50 issues per run) and Microsoft Planner (filter tasks by bucket). Pipedrive gained 13 actions around projects and tasks; they require the additional projects:full scope, so existing connections need to be re-authorized once.

14.2 Establishing Connections

For most integrations OAuth is enough: click »Connect«, sign in with the third-party, approve the requested permissions, done.

• Personal connection: You are connected with your own account, only you use the integration in your name.
• Workspace connection: Set up by the admin, used by everyone. Actions run under a central service account.

The choice depends on the use case: when employees should search their own emails, personal. When an agent works for the whole team in a central Salesforce account, workspace.

Since August 2026 connecting is tidier in several places. Admins can register multiple custom OAuth clients per integration and assign them to specific groups, for example separate connections for two country organisations. Integrations that need an API key instead of OAuth now show, for more than 20 services, exactly where to obtain that key in the target system. Public integrations carry a risk level in the directory, which makes the approval decision easier for admins. And actions that send or write something show a sanitised preview of the content before confirmation, so email and Markdown inputs appear as finished text rather than raw data.

14.3 Three Paths to SharePoint and Drive Data

Three mechanisms exist to make files from SharePoint or Google Drive addressable. Each solves a different problem.

The Folder Sync limits that bite first in practice: an agent or project takes at most five synced folders. Per folder, the cloud processes 200 files, dedicated deployments 1,000. Anything beyond that is silently skipped, in the order the source system returns the files. Subfolders count toward the limit, so attaching them individually as their own synced folders is often worth it. Images are excluded, text documents go through semantic search, and spreadsheets are read directly. For Confluence Cloud, the sync imports the pages of a named folder, not the attachments hanging off those pages.

14.4 Microsoft Integrations: Permissions & Admin Approval

Microsoft integrations (SharePoint, OneDrive, Teams, Outlook, Entra) often require admin approval before users can connect them. Background: M365 tenants have default policies that need to allow OAuth apps first.

• Who approves: Tenant admin (Entra Global Admin or delegated).
• What gets approved: The Langdock app registration in your tenant.
• Where it stalls: User sees a »need admin approval« message at connect time. Admin finds the request in Entra under »Enterprise Applications → Admin consent requests«.

Details and step-by-step screenshots in the Langdock docs under Microsoft Integrations: Permissions & Admin Approval.

14.5 SharePoint Page Links (.aspx)

Since May 2026, Langdock automatically recognizes SharePoint page links (.aspx) as attachments. Whether you paste the link into chat or forward it to an agent, the page content is read like a regular file. Useful for wikis and intranet pages that previously needed Folder Sync or an Action.

Direct paste in the file picker: URLs to supported sources can be pasted directly in the file picker, skipping the Integration sidebar detour.

14.6 Reauthorize Flow

When an integration's OAuth token expires, Langdock now shows a reauthorize prompt directly in chat (since May 2026). One click reauthorizes and the chat continues. Admin Settings → Integrations adds clearer alerts for workspace-wide connections. Since June 2026 you can also re-authenticate an integration straight from settings, without having to remove it and set it up from scratch first.

14.7 Langdock for Excel (Plugin)

Since early July 2026 an Excel plugin brings Langdock straight into the spreadsheet. You work in your familiar workbook and let Langdock read, write and create new files without leaving the application. The plugin taps into your existing integrations, knowledge bases, skills and workflows.

• What it does: answer questions about the workbook and spot anomalies, generate formulas, tables, charts and pivot tables from plain language, clean up messy data (standardize names, dates, categories, currencies), name trends and outliers, and draft summaries for reports or presentations.
• Context access: from inside the plugin you reach your knowledge bases, connected integrations, web search, skills and workflows, just like in chat.
• Library: finished spreadsheets can be saved straight to the Langdock Library and shared with the team.
• Install: via the Microsoft Marketplace (»Get it now« → »Open in Excel« → sign in with your Langdock account).
• Admin enablement: a workspace admin has to enable the plugin in the integration settings first, only then can users install it.

14.8 Langdock for Outlook (Add-in)

Since late July 2026 the same principle applies to Outlook: an add-in from the Microsoft Marketplace that puts Langdock in a side pane next to your inbox and calendar. Langdock sees which email or calendar event you have open, so you ask about it directly instead of copying content into a separate chat. Available in Outlook on the web and on the desktop.

• What it does: summarise and prioritise your inbox, boil long threads down to key points, open questions and next steps, draft replies and refine their tone, compose new emails including subject and recipients, sort messages into folders, check your schedule and create or reschedule events.
• Context: models, knowledge bases, integrations, web search, Skills and Workflows are the same ones you have in Chat.
• Installation: via the Microsoft Marketplace (»Get it now« → »Open in Outlook Web«), then open an email in Outlook, click »Langdock« in the toolbar and sign in with the code from your browser. Without an open email the toolbar apps stay disabled.
• Two integrations required: the add-in uses Outlook Email and Outlook Calendar to read messages, search your mailbox and work with your calendar. Connect both with the same Microsoft account you use in Outlook. Without them, Langdock only knows the metadata of the open item (subject, participants, timing, location, attachment names) and whatever you have typed into an open draft.
• Pin the pane: Outlook closes an unpinned pane when you switch between emails. Click the pin icon at the top of the Langdock pane once, and once more in a compose window.
• Edit permissions: the default is »Ask before edits«. Before Langdock writes into an open draft, a dialog shows the fields it is about to change, with allow once, always allow or deny. You can switch to »Accept all edits« in the chat input menu. In both modes, edits only ever touch the open draft, and nothing is sent unless you ask for it and confirm.
• Your own preferences: the add-in menu under »Preferences« lets you set custom instructions that apply only to add-in chats, plus a default model, independent of your global account settings. Theme and deleting all add-in conversations live there too.
• Admin enablement: as with the Excel plugin, a workspace admin enables the add-in in the integration settings first. Outlook Email and Outlook Calendar have to be active in the workspace as well.

14.9 Scopes & Action Access (Admin)

Every integration action needs specific OAuth scopes from the provider, an Outlook Calendar action for instance Calendars.ReadWrite before it can create events. Under workspace settings → Integrations, the Scopes and Action Access area shows per integration which scopes Langdock requests and which actions that makes usable. This is where you answer the question »why does this action say Missing scope«.

• Langdock client (the cloud default): Langdock keeps requested scopes in sync with the actions that are enabled or shared. You don't edit scopes by hand, you steer them by enabling, disabling, or sharing actions.
• Own OAuth client with scope sync: behaves the same way, just with your own app. The recommended setup when you run your own client but don't want to maintain scopes yourself.
• Own OAuth client without scope sync: a fixed scope list you manage manually. The route for tenants with a strictly approved scope set.
• Two views: the scope view marks with a green check or a red cross which scope groups are currently requested. The action view goes action by action, with search, sorting by state, and the scopes each action needs.
• New actions enabled by default: controls what happens when Langdock adds an action to an existing integration. Turned on, actions whose scopes the client already covers are enabled automatically and the rest stay off. Turned off, every new action stays off until you review it. Either way, a new scope never silently expands what Langdock may request from your provider.
• After a change: when a scope is added or removed you confirm it in a dialog. Users then have to refresh their connection once before the new scope actually takes effect.

Several custom OAuth clients per integration are possible: one for the whole workspace as the default, plus others assigned to specific groups, for example a subsidiary with its own Entra tenant. Existing connections stay bound to the client they were created with, only new connections move to the activated client.

Skills are reusable instructions that activate automatically in chat once the context fits. Unlike prompts you don't need to invoke them.

15.1 Skills Concept

A skill is a piece of instruction with a trigger condition: »When the conversation is about topic X, apply this rule.« Examples:

• Recap skill: when someone asks for a meeting summary, always reply with »attendees, decisions, to-dos«.
• Translation skill: on translation requests, always keep second-person tone.
• Compliance skill: on legal topics, always add the »no legal advice« disclaimer.

15.2 System Skills

Built-in, always active for all users. No installation, no configuration, can't be disabled. There are currently seven system skills, maintained centrally by Langdock:

• File generation (four skills): PDF, DOCX, XLSX and PPTX. They run invisibly in the background whenever a file is created or read in chat, and never show up in the skills list.
• Data visualisation: Creates charts and visualisations from data. Also runs in the background without its own entry in the list.
• Skill Creator: Walks you through creating and updating your own skills in conversation (see 15.4). Visible on the skills page.
• Platform Help: Added in August 2026. Answers questions about using Langdock right inside chat, with an eye on your own workspace. Also visible on the skills page.

System skills count neither towards the personal limit of 30 active skills nor towards the eight mandated workspace skills (see 15.3 and 15.4). File generation does not depend on the skills infrastructure and stays available even when skills are switched off in the workspace.

15.3 Workspace Skills

Only admins can roll a skill out to the whole workspace. It is controlled in the skill's share dialog under »General access«, and there are two levels:

• Shared workspace-wide: »Anyone in workspace«. The skill is discoverable for everyone, but only becomes active once someone switches it on themselves. Good for offers that not every role needs.
• Mandated: additionally »Apply for all members«. The skill is active for everyone automatically, the toggle is locked, nobody can opt out. Useful for compliance instructions (e.g. »no PII in answers«).

Since the plugins rework in mid-September 2026, skills have a single state: enabled or disabled. The previously separate install step is gone, two steps became one toggle.

A workspace supports a maximum of eight mandated skills at a time. They do not count towards the users' personal limit. Members and editors can create skills and share them with individual people or groups, but cannot roll them out workspace-wide.

15.4 Creating Your Own Skills

Three ways:

1. Manual: Write instructions in the UI, define trigger conditions.
2. Via chat generator: »I need a skill that ...«. The model proposes a configuration, you confirm.
3. Upload a SKILL.md file, ZIP bundle, or a .skill file. For skills you developed elsewhere or want to import. Since July 2026, skill packages up to 20 MB uncompressed are allowed (previously 10 MB), and the ZIP itself may be up to 25 MB. Enough room for skills that ship several bundled resources.

Since May 2026 skill creation is smarter: Langdock detects duplicates and updates existing skills instead of creating new ones every time. Anyone iterating on a skill avoids the proliferation of half-baked variants.

Edit in place: Since August 2026 you change an existing skill where it lives, no re-upload needed. Name and icon via the header, description and instructions in the SKILL.md editor, plus adding, editing, renaming, moving or deleting supporting files and adjusting the linked integrations. SKILL.md itself stays the main file and cannot be renamed, moved or deleted.

Limit: Each user can have 30 active skills at a time. That is not a cost cap but a quality boundary: the more skills are active, the more often the model reaches for the wrong one. If you have collected a lot of them, deactivate the ones you do not currently need.

Allowed from member role upward. Sharing same as for prompts (users, groups, workspace). Since June 2026 you get notified as soon as someone shares a skill with you, so you no longer miss a new template.

15.5 Coupling Skills With Integrations

Skills can call integrations. Example: a »sales lead skill« detects a lead query in chat, automatically pulls Salesforce, and enriches the answer.

Prerequisite: the integration is connected in the workspace and the user has access.

15.6 Skill vs. Agent vs. Prompt: When What

Agents are specialized bots for recurring tasks. With their own knowledge, tools, and instructions. When a use case is common enough that everyone should handle it the same way, it belongs in an agent.

16.1 Internal vs. External Agents

16.2 Agent Configuration

Two ways to build an agent. Since July 2026 you can build an agent conversationally: under »Agents → Create agent« you describe in plain language what the agent should do (e.g. »An agent that answers onboarding questions for new employees based on our handbook«). The builder asks targeted clarifying questions with clickable answer options, suggests matching integrations, knowledge bases and existing agents from your workspace, and configures things live while you watch. The builder chat stays available in a side panel, and »Build« and »Test« switch between configuration and a trial chat. Editing fields manually and the dialog both act on the same configuration. Whether via the dialog or by hand, you set the following building blocks.

• Name & description: What does the agent do? (Users see this in the picker.) Name up to 80 characters, description up to 800. Markdown is rendered in descriptions: headings, bold, italic, code, links.
• Model: Recommended: one matching model per agent (routine bot → Haiku, strategy bot → Opus). See chapter 3.2. You also set what happens at the usage limit: switch model (the default, the agent keeps running) or block usage, for agents that only deliver reliable results on the chosen model.
• Creativity: A temperature slider from 0 (deterministic, the same question gives practically the same answer) to 1 (varied). New agents start at 0.7, or 0.3 once knowledge sources or tools are attached. Keep it low for lookup and classification agents, higher for ideation and writing agents.
• System prompt: The central instruction. Who is the agent, what does it do, what does it not do. Since June 2026 up to 50,000 characters are allowed, plenty of room for detailed rules, examples, and format specs.
• Knowledge: Assign one or more folders (chapter 12). An agent holds up to 50 files, plus 5 Library folders and 5 synced folders. The 50 files arrived in July 2026, up from 20.
• Tools: Activate integrations the agent is allowed to use. With the latest models (OpenAI from GPT-5.4, Anthropic Sonnet/Opus 4 and newer), since June 2026 the model natively picks the right tool out of many enabled ones. So you can give an agent plenty of actions without hurting hit rate. Since July 2026 this automatic tool search can be opted into per agent, and the AI can use up to 15 results from an integration search instead of just a handful. If you attach a workflow, skill or nested agent that the intended users or groups cannot access, Langdock warns you while configuring.
• Skills: Attach one or more skills (chapter 15) that fire automatically in every agent chat. Available for projects too.
• File templates: Since May 2026 templates can be selected directly on the agent so every run lands in the right format (chapter 13.2).
• Conversation starters: Suggested first prompts users can launch with one click.
• Creator display name: Since July 2026, workspace admins can set a custom creator display name per agent in the agent settings. It shows up on agent cards, the chat start screen and the agent info dialogs. Useful when a whole team built the agent and a single person should not appear as the author.
• Restrict source access: Optionally the agent shows its citations without granting access to the underlying file. For cases where the answer may be shared but the source document may not. A workspace admin enables the option first.
• Tracing: Langfuse can be connected optionally to record runs in detail. An admin enables agent logging in the workspace, then you turn on the agent-logs toggle on the agent and set the tracing URL. Relevant for teams that debug agents systematically or want to measure quality over time.
• Working folder: Since September 2026 you can already decide in the Agent Builder chat which folder the agent works in. No need to finish the agent first and add the folder in the configuration afterwards.
• Form fields or free text: How users interact with the agent (see 16.3).

The editor was reorganized in August 2026. Instead of one long form there are clearly separated sections, each with its own search. Instructions sit at the top in an expanded editor, which is noticeably more comfortable for long rule sets. Knowledge has dedicated controls for files and for folders. Tools are grouped into Integrations, Skills and More, and selecting an integration pulls in its actions in one step rather than one by one.

The real addition is per-action approval: every tool can be set to Always allow or Needs approval. Read actions run straight through while write actions ask first. “Apply to all actions” sets the choice for an entire integration in one go. A connection can also be pre-selected for all users, so nobody has to assign one on first use.

16.3 Form Fields

Instead of free-text input the agent gets structured input through defined fields. Guarantees consistent inputs and less variance in outputs.

There are nine field types:

Every field has a label (required), an optional explanatory description and the »required« toggle. Multi-select arrived in August 2026 and fits wherever several options apply at once: affected departments, product categories, required report sections. If the field is required, at least one option must be picked. Example for a »letter generator« agent:

• Recipient (text)
• Reason (select: reminder, offer, reply, other)
• Tone (select: formal, friendly, brief)
• Mandatory elements (multi-select: deadline, contact person, list of enclosures)
• Context (multi-line text)

Boundaries: a maximum of 25 fields per form when you build it yourself, a maximum of 20 when the Agent Builder generates it. Labels and option texts hold 255 characters each, descriptions 512.

16.4 Subagents

An agent can call other agents. Useful for multi-stage tasks where each stage needs a different specialist.

Example »content pipeline«: master agent → research agent (collects) → writer agent (drafts) → review agent (checks style and facts) → final output.

Advantages: each subagent has an isolated, smaller context. Saves tokens compared to one long chat that drags everything along. Better quality through focus.

Since June 2026 subagents can run significantly more steps per run, so longer multi-step tasks complete more reliably. Subagents can now also be addressed as tools via the API, letting you build nested agent setups programmatically. Since August 2026 a visual version history shows how a subagent has changed over time, so in multi-stage setups you can see which stage caused a shift in behaviour.

The boundaries remain: subagents cannot call subagents of their own, they do not see the parent agent's conversation, and they cannot ask the user clarifying questions. Only the parent agent does that.

16.5 Agent Templates

Langdock ships a library of pre-built agent templates, from sales outreach through marketing reviewer to procurement-request reviewer. When creating a new agent you can use a template as a starting point and adapt it.

16.6 Versioning & Draft Mode

Since April 2026 agents have real versioning. Changes land in a private draft that auto-saves while you edit. Things only go live when you explicitly hit »Update« in the top right.

• Test draft: verify in the editor whether the change behaves as intended. Only you see the draft.
• Publish summary: on update, Langdock shows what changed compared to the live version (e.g. »name updated, instructions updated«). Optionally add a description.
• Version history: all versions remain. Rollback to an earlier one is possible at any time.

16.7 Usage Insights

Since April 2026 every agent has its own Analytics tab at the top of the configuration. There you see:

• Number of conversations per day/week/month
• Active users: who uses the agent regularly
• Top conversations: what are the typical use cases
• Adoption funnel: how many of the people with access move from access to regular usage, plus the list of the most active users
• Most used actions: which integrations the agent actually pulls
• Cost: cost per message, once enough data has accumulated
• Feedback: thumbs reactions plus filters, conversation switcher and preview directly inside the analytics tab. You don't have to leave the page for reviews.
• Filter feedback by version: since May 2026 thumbs reactions can be scoped to a specific agent version. That tells you whether a new publish measurably improved or hurt quality.
• Feedback nudge: since May 2026 users are gently invited to leave a thumbs-up or thumbs-down. Since June 2026 the prompt appears automatically once someone has had more than five conversations with the agent. That fills the analytics tab with enough signal without owners having to chase people.
• Shareable feedback link: since June 2026 you can generate a feedback link from the menu in the chat header and pass it to colleagues. That collects targeted feedback, even from people who rarely use the agent themselves.

Time range and export: the last 30 days are preselected, alongside a »Last 24 hours« preset and a freely configurable range. Analytics and feedback can be downloaded separately as CSV. On feedback itself, users can optionally add name and email next to thumbs and comment, and share the entire chat, which only the agent's creator gets to see. Both tabs are visible to creators and editors of the agent.

Valuable for agent owners: when adoption stays low after 4 weeks, it's usually not the model's fault. The agent doesn't fit the actual use case.

16.8 Advanced Features

• Labels: tagging for better sorting in the agent overview (e.g. »sales«, »HR«, »compliance«).
• Pinning: pin favorite agents in the left sidebar.
• Duplication: copy an existing agent as a starting point for a variant.
• Owner transfer: hand the agent to someone else (important on team changes).

16.9 Slack & Teams Bots

Agents can be exposed as bots in Slack or Microsoft Teams. One-time setup per workspace, then all approved agents are reachable via @-mention in channels or DM.

• Slack setup: Workspace settings → Integrations → Slack → install bot. Pick channels where the bot may respond.
• Teams setup: Same for Microsoft Teams.
• Threads in Slack/Teams become conversations in the agent. Context stays inside the thread.
• Tip for internal channels: Set up a #ai-help channel where a help bot lives. Lowers the barrier for asking questions.

16.10 Sharing & Permissions

Roles: Reader (use only), Editor (change configuration), Owner (everything including delete).

If someone hits a private agent they can't use, there is a »Request access« button (since June 2026). The request goes straight to the owner, who can grant access in one click, no need to chase down the link by other means.

Since August 2026 permissions are drawn more tightly: only users explicitly granted the right can manage access to an agent. Governance admins can additionally handle ownership and sharing of private agents, which makes orphaned agents resolvable. When a user is deactivated via SCIM, their agents show up as unassigned instead of quietly staying with the departed account.

16.11 Agent Evals: test before you publish

Since late June 2026 you can test an agent with structured test sets before publishing a new version, right in the agent editor via the Agent Evals tab (with the »Test sets« and »Runs« sections). That way you catch issues before they reach your team.

• Test sets: a test set bundles multiple test cases (prompt plus expected answer) under a shared configuration. Add cases one by one via »Add case« or import them via CSV to build larger sets quickly.
• Three grading methods: AI judge (a judge model compares the response semantically with the expectation, not word for word), Tool check (verifies the agent invokes the right tools) and Keyword check (»Must mention« / »Must not mention«, deterministic, no model). With the AI judge you pick the evaluating model yourself.
• Tool execution: the default is Dry run, calls are only recorded, not executed. In Live mode real actions run (except those requiring approval), useful for end-to-end tests, but it can change external systems (emails, tickets, records).
• Run & results: »Run« starts the evaluation, results stream in live per case (up to 100 cases per run, only one active run at a time). You see the transcript, token usage, duration and each grader outcome, and can export everything via »Download CSV«.

16.12 Archiving Agents

Since September 2026 agents can be archived instead of deleted. Archiving takes an agent out of the active list without removing it: the configuration is preserved and you can bring it back at any time. This is the remedy for agent sprawl in workspaces that have grown over time.

• How it works: Owners archive the agents they own. The Archived tab is where you find them again, and Restore agent brings one back.
• Admin route: Workspace admins can archive any agent from the governance area, with a reason and an optional note. Details in ch. 20.4.
• Automatically: Unused agents are archived on their own after a window the admin sets. If you still need an agent, simply use it during the grace period and the scheduled archival is cancelled.

Workflows are multi-step automations with logic, conditions, external calls. Unlike agents they are deterministic. Same input gives same output. And they do not count against the session or weekly limit.

17.1 Concept & Triggers

A workflow starts with a trigger and ends with an output. In between, nodes are chained:

• Manual trigger: start by button click in the UI
• Form trigger: user fills a form, workflow runs
• Webhook trigger: external system calls the workflow
• Scheduled trigger: time-based (cron-style)
• Integration trigger: event in a connected system (new email, new lead, new calendar entry)

17.2 Node Overview

18 node types for different functions:

HTTP Request has three auth modes: no auth, header auth (recommended), query-param auth (legacy).

Webhook trigger supports three auth variants: no auth, query parameters and an X-Webhook-Secret header. For production setups: use the header variant. When »Wait for response« is enabled, a »Respond to Webhook« node is automatically appended.

MCP server connections in workflows: HTTPS URLs with query parameters are accepted. Workspace-shared MCP connections work for all members.

Extended thinking on agent nodes: Since May 2026 reasoning mode can be enabled per agent node. Worth it for steps with real cognitive load (complex classification, strategic decisions in the flow), keep it off for routine steps.

Richer file metadata: Downstream nodes now also get file ID, external ID, connection ID, size and page count as variables. Useful for routing, logging and cost tracking.

17.3 Variables, Field Modes, Cost Cap

• Variables: Outputs of one node are available as variables in subsequent nodes ({{node1.result}}).
• Field modes: Fields in nodes have three modes: static (fixed value), dynamic (from variable), AI-generated (model fills at runtime).
• Cost cap: Set a budget limit per workflow. On overrun, the run is aborted.
• Human in the loop: Approval gate as a node. A human approves before the workflow continues.
• Model on the output node: Since August 2026 output nodes set to »Prompt AI« or »Auto« come with a model picker, including a warning when the choice does not fit the task. Structured inputs now accept video uploads as well.

17.4 Workflows From Chat via @-Mention

Since Q1 2026 workflows can be invoked like agents directly from any chat via @. Useful for workflows that should accept ad-hoc requests without a separate UI. Since July 2026 this workflow chat also supports Opus and reasoning models including the extended-thinking toggle, so more demanding requests can run here at full thinking power.

Public workflow forms can be embedded as iframes in your own websites or internal portals. A workflow becomes addressable outside the Langdock UI without API code.

Default limit: 100 workflow executions per hour (up from 30). Workspace limit can be raised for higher-volume use cases, up to a maximum of 5,000 per hour according to the limits docs. On top of that, runs default to up to 2,000 steps, and Cloud workspaces have a monthly cap of 2,500 runs (depending on plan).

17.5 Composing Workflows

Since 18 September 2026 workflows can be combined into larger processes instead of building every automation as one long standalone workflow.

• Workflows call workflows: In the editor you select another workflow, map its inputs and use its result in the steps that follow. This works up to five levels deep, and Langdock detects cycles automatically. Build recurring building blocks (say »enrich lead«) once and reuse them everywhere.
• Multiple triggers per workflow: The same automation can start from different sources, for example from a form and on a schedule.
• Workflow Monitor: Flags invalid connections before a run fails on them and reruns up to 20 failed runs at once.
• Further additions: environment variables, workflow run exports, imports from Zapier and Make, file access in the code node, an audio transcription node, webhook triggers that also accept GET requests, test runs that keep their context, and a redesigned display of output values.

17.6 Example Patterns

Three recurring patterns that are useful in almost any company:

• Approval: Form trigger (e.g. vacation request) → agent assessment → human approval → output (approval / rejection with reasoning).
• Enrichment: Webhook trigger (new lead from form) → web search for the company → agent classification (ICP fit) → CRM update via action.
• Scheduled report: Scheduled trigger (Monday 8am) → data aggregation across integrations → agent synthesis as Markdown → send notification to a Slack channel.

When the native integrations aren't enough, because the system is an internal build or a very specific action is needed, there are two paths: custom integrations and MCP.

18.1 Custom Integrations & Action Builder

Define your own action that calls a REST API. Per action: endpoint, auth, request schema, response schema.

Action Builder Agent: Helps you build by letting you describe in natural language what the action should do. The agent generates the configuration and you confirm.

Integration Builder Agent: Builds the whole integration from a conversation: authentication (OAuth, API key or none), actions and triggers. You describe the tool and what you need, attach the API docs if you have them, the agent researches the rest itself and ends with a summary including auth type and scopes. Prerequisite: a connection to the Langdock integration with an API key that carries the Integrations API scope (a workspace key from your admin, or a personal key if your admin allows that scope). Start it via Integrations → Add integration → Start building, or as a template under Agents → Browse templates → Duplicate Agent. On dedicated deployments the banner is missing, so only the template route works there. Afterwards you finish authentication on the Integrations page.

18.2 Model Context Protocol (MCP) as a Concept

MCP is an open protocol for communication between AI models and external tools/data. Similar to REST for web services but standardized for AI tools. Providers like Anthropic, OpenAI, and Langdock support MCP.

Advantage: anyone who builds an MCP server can connect it to all MCP-compatible clients. No custom code per integration.

18.3 Connecting Custom MCP Servers

Connect your own MCP server (self-hosted) to Langdock via URL and auth. The server's tools then become available to agents like native integrations.

Available in the integration directory as »Add MCP server«. Auth typically via API key or OAuth.

Four steps to connect: enter the server URL, pick an auth method, hit »Test connection« to see which tools the server exposes, then select the tools you want and save them. Langdock supports STREAMABLE_HTTP and SSE as transports. Private and internal network addresses are blocked by default.

Four auth methods are available:

• No auth: for public servers that don't require identification.
• API key: the header type is selectable: X-API-Key, API-Key, X-Auth-Token, Authorization: Bearer, Authorization: Token or Authorization: Basic (base64-encoded, in which case enter the key as user:password). Without a selection, the legacy variant applies and sends the key both as X-API-Key and as a bearer token.
• OAuth: the standard path for servers with PKCE and Dynamic Client Registration: enter the URL, click »+ Add connection«, run through the OAuth flow. Langdock discovers supported metadata and scopes automatically where the server provides them.
• Advanced OAuth: for servers without Dynamic Client Registration. Store Langdock's redirect URL in the target application first, then enter client ID, client secret, authorization URL, token URL and the required scopes yourself. Leave »Send resource parameter« enabled in most cases; switch it off for Entra ID, where the resource is part of the scope.

Limits and defaults: up to 60 tools and 50 resources can be enabled per integration. Newly saved tools and resources run without user confirmation by default; confirmation can be switched on per tool afterwards. Custom headers accept static values plus the placeholders {{ access_token }}, {{ refresh_token }} and {{ api_key }}.

MCP directory: ready-made servers instead of self-hosting. Before you host anything yourself, check the MCP directory. It currently lists 79 official, remotely hosted servers (as of 15.10.2026), maintained by the vendors of the respective products, among them Amplitude, Asana, Atlassian, Canva, ClickUp, GitHub, HubSpot, Linear, monday.com, Notion, Stripe, Vercel and Zapier, plus infrastructure services such as Supabase, Neon, Cloudflare Workers and Netlify and payment providers such as Square and PayPal. Every entry states the server URL, the auth type (none, OAuth 2.0, API key or custom), the transport (HTTP or SSE) and a link to the vendor docs. No local installation, you only enter the URL and credentials. Langdock only lists servers it has tested itself; vendors apply for inclusion via support@langdock.com, with server URL, auth type, docs link and a test account.

Shared MCP connections: Since May 2026, MCP connections can be shared workspace-wide instead of every user reconfiguring the same server. Admins set up the connection once, and the tools are available to the whole team.

Passing through user tokens: Since July 2026 Langdock understands the placeholder {{ user_oauth_access_token }} in MCP server configuration. The server then calls an action with the token of the requesting user instead of a shared service account, which is what you want wherever individual permissions should apply. At the same time, Langdock only reads MCP resources from URI actions that are explicitly enabled, and whether the resource parameter is sent for manual MCP OAuth is now configurable.

18.4 Langdock Agent MCP Server

The reverse also works: Langdock can act as an MCP server itself and expose workspace agents as MCP tools to external clients.

• Use case 1: Claude Code (Anthropic) calls your internal sales agent as a tool.
• Use case 2: A custom application using the OpenAI SDK uses your HR agent.
• Use case 3: Other AI platforms with MCP support hook into Langdock agents.

18.5 MCP Apps: Interactive Interfaces in Chat

An MCP server normally returns text. With MCP Apps, an official extension to the protocol, it can return a small interactive interface instead: a form, a dashboard, a visualisation. It renders directly in the chat, and on desktop clients it can be moved into a resizable side panel while the conversation continues.

• How it works technically: The tool points to a ui:// resource via _meta.ui.resourceUri. Langdock fetches the HTML and renders it in a sandboxed iframe. The interface talks to the host through JSON-RPC over postMessage, so it can request tool calls and update context.
• Building one: Through the @modelcontextprotocol/ext-apps SDK, compatible with React, Vue, Svelte, Solid, Preact and vanilla JavaScript.
• Security: Sandboxed iframe, the host reviews the HTML before rendering, all app-to-host communication is loggable, and tool calls still require user approval.
• Compatibility: Besides Langdock, Claude Desktop and VS Code support the standard. MCP servers without a UI resource keep working unchanged, with plain text output.

18.6 A2A: Connecting Agents to Other Agents

MCP connects agents to tools (databases, APIs, services). The A2A protocol (agent-to-agent) connects agents to other agents, across platform boundaries: delegating instead of doing it yourself. The open standard originated at Google and is now governed by the Linux Foundation. Both can be combined in the same setup.

• Discovery via the AgentCard: An agent publishes a JSON file at /.well-known/agent-card.json listing its name, description, URL, version and skills. That is how agents discover each other, without anyone documenting the interface by hand.
• Four-step flow: Discover capabilities, submit the task with its input data, the remote agent processes it, the complete result comes back.
• Current limits: No streaming, A2A waits for the complete response. Authentication is limited to none or API key. Langdock blocks private and internal network addresses by default, which can be opened up via URL_VALIDATION_ENABLED=false.
• Which when: A2A for delegating specialist tasks and coordinating several agents. MCP for database queries, API calls and connecting external tools.

Langdock has an API for anyone who wants to embed AI capabilities into their own applications. Three entry points, depending on the use case.

19.1 Facade Mode

Existing SDKs from Anthropic, OpenAI, Google, or Mistral simply point to Langdock. No code change, just base URL and API key. Benefits: workspace limits, EU hosting, central billing apply to these calls too.

19.2 Native API

For functionality specific to Langdock:

• Agents via API including their knowledge connection and tools. Since May 2026 with a supportsExtendedThinking field to toggle reasoning per call.
• Folders management (create, upload files, search)
• Audit logs retrieval (for compliance reporting)
• Usage export (per model, agent, project, user)
• Manage scheduled tasks. Since mid-September 2026 scheduled tasks can be created, listed, updated, paused, resumed, run immediately and deleted through the API (see chapter 4.12).
• Manage workflows. The Workflow API lists, creates, updates, publishes and deletes workflows and reads run history page by page. Three separate API key permissions control reading, writing and deleting. Newly created workflows stay private until they are shared.
• Manage the prompt library. Since August 2026 /api/public/prompts/v1 offers endpoints to create, read, update and delete prompts, with cursor pagination for large libraries. Useful for keeping a curated prompt collection in sync from another system.

19.3 Rate Limits & Auth

• Rate limits: 150,000 tokens/minute, 500 requests/minute (requests per workspace, model and API key; tokens shared across all keys using the same model; higher limits via »Model quotas« in the workspace API settings). Usage export endpoints capped at 1M rows per call.
• Auth: API keys are created in account settings. Per key you can restrict which agents/folders are accessible.

Personal API keys for coding tools: A personal key lets you use your workspace's models from the outside, for example in a development environment or your own project. The key is tied to you personally, reaches the completion endpoints only, and draws on the same budget as your chats and agents.

• The key is you. It authenticates as your person and spends your budget. Never share it; shared applications and automations call for a workspace key instead.
• Admins enable it. Access is granted under Settings → Workspace → Products → API, either workspace-wide or for specific members and groups. A usage limit above zero has to be set as well.
• No fallback at the limit. Once the budget is used up, the API returns an error instead of falling back to a cheaper model.
• Admin view: monthly spend, number of active keys and a 14-day per-member trend. Individual keys can be inspected and revoked immediately.
• Base URL: api.langdock.com with a region segment in the path. On dedicated deployments, <deployment-url>/api/public replaces api.langdock.com.

What admins need regularly, kept compact. For SAML/SCIM/BYOK setup see chapter 21.

20.1 Plans & Subscription

Billed per seat, monthly or annually. Invoice history in workspace settings.

20.2 Extra Usage

Workspace settings → Usage → »Enable extra usage«. Default cap is €1,000/month workspace-wide. Anything above is billed at API tariffs plus surcharge as a separate line item on the monthly invoice.

• Recommendation when activating: Don't stay at the default cap. Lower it to €200 to 300 first, recalibrate after 4 weeks based on usage exports.
• Per-user override: Set an individual budget per user (Standard, Unlimited, or custom amount). Useful to handle heavy users without blowing the workspace cap.
• Allow usage requests: Keep on. Users can request a limit increase directly from chat. The admin sees it in the inbox plus by email.

Exporting API costs: since April 2026 API costs can be exported as CSV over a freely chosen period of up to 12 months. More practical for FinOps reviews and cost-centre chargebacks than screenshots from the charts.

Usage credits (prepaid balance instead of invoicing later): some workspaces no longer bill extra usage, API usage and model usage in workflows at the end of the month but deduct it from a prepaid balance. Whether that applies to your workspace is shown under Workspace settings → Usage & spend: if there is a »Usage credits« section, an admin must buy credits first (»Buy credits«, by card, Apple Pay, Google Pay or PayPal, amounts excluding VAT) before this usage can be enabled at all. If the section is missing, nothing changes. »View credit history« lists all purchases, auto-reloads and adjustments.

• Turn on auto-reload: Langdock automatically tops up the configured amount before the balance runs out. If a payment fails, admins get an email and add a new payment method under »Update payment method«; the reload then resumes on its own.
• When the balance is empty: without auto-reload, admins are warned by email at €10 and at €0. Usage included in the plan keeps working, but extra usage, API and model usage in workflows pause until more credits are bought. Anyone who has already hit their limit drops to the fallback model. For workflows running in production, auto-reload is therefore a must.

Agent limits: since August 2026 an agent can have its own monthly limit, separate from user limits. Useful for business-critical agents that shouldn't stop when someone runs out of their personal limit, for expensive agents whose cost you want to track separately, and for shared team agents that need a predictable ceiling. Find it under workspace settings → Usage → the Agents tab, then »Edit limit« per agent with a monthly amount or unlimited. Left at the default, an agent has no limit of its own and its usage counts against personal limits as before.

• After user limits (default): usage counts against personal limits first. Once someone reaches theirs, further usage counts against the agent's limit. The safety net that only steps in when a budget runs out.
• Always: usage counts against the agent's limit immediately and never touches personal budgets. The mode for a shared agent that shouldn't consume individual budgets.

You can only pick the mode once a limit is set. Setting it requires workspace admin access, while the agent owner decides on the edit page what happens when the limit is reached. In workspaces with user-level data disabled, private agents aren't listed and have to be shared first.

Spend limits in BYOK workspaces: when a workspace runs on its own provider keys (see ch. 21.3), usage is capped by monthly spend rather than by messages. Limits are measured against actual provider cost and managed under workspace settings → Usage. Three scopes combine: a workspace spend limit for total spend, per-user limits from plans or groups, and agent limits as described above.

• Two plans: BYOK workspaces have Standard (the default) and Power user. Move people who need more headroom in the Plan column on the Users tab.
• Weekly and session windows: Langdock breaks every monthly limit down further so a single busy day can't burn the month. A week gets a quarter of the monthly amount, a session a third of the week over a five-hour window. If your team needs bursts, expand »Weekly and session limits« and pick Use weekly limits only, which then applies workspace-wide.
• Group limits: on the Groups tab a team gets its own monthly budget. When more than one rule applies to someone, the highest limit across plan and groups wins. For SCIM-provisioned groups the limit stays attached to the group, so new members inherit it automatically.
• Extra usage per person: a flat monthly amount on top, with no weekly or session breakdown. It only kicks in once the session or weekly limit is exhausted. Set it on the Users tab via »Edit individual limit«.
• Fallback model: once a limit is reached and no extra usage remains, the next message continues on the cost limit fallback model, configured under workspace settings → Models. Without an explicit choice, the workspace backbone model is used. Pick something cheap but capable, so people keep working instead of stopping.
• Personal API keys: their usage counts against the workspace spend limit and individual extra usage, but not against plan or group limits. When either is exhausted, the API returns an error instead of switching to the fallback model. Backbone usage, workflow runs, and workspace API keys aren't counted at all.
• Models without pricing: cost tracking needs input and output token prices on the model. Without them the usage page shows a warning and usage on that model counts against no limit. So always add prices when you bring in your own models.

Until per-user limits are set, users see nothing of this. After that they track their own position under account settings → Usage and in the context window indicator next to the model selector. If a limit is reached while a model is responding, it finishes that response, and only the next message moves to the fallback model.

20.3 User Management & Permissions

• SCIM for auto-provisioning (Microsoft Entra, Okta via generic SCIM endpoint). When someone is deactivated via SCIM, their scheduled tasks are preserved as of August 2026 and resume on reactivation instead of being lost.
• Roles: Member (default), Admin (workspace settings), Owner (billing).
• Permission Recommendations in the docs. Good defaults for typical setups.

20.4 Manage Agents & Manage Workflows

Both areas live in the admin menu, with the same structure:

• Verified / highlighted: Mark vetted agents/workflows so they appear at the top of the picker.
• Disabled: Temporarily switch off without deleting.
• Owner transfer: On team changes, otherwise orphaned agents/workflows pile up.
• Spend limits for workflows: per-run cap to prevent escalation.
• Security controls: Which actions a workflow may run (e.g. no external HTTP requests).

Plus: Admin Integration Newsletter, an opt-in notification system for critical integration updates that require admin action.

Managing scheduled tasks: Once »Scheduled« is active in the workspace, admins get their own overview of all scheduled tasks, with name, owner, frequency, status, number of runs in 30 days and last execution, plus search, filters and sorting. The central admin action is the on-off switch: »Turn off« stops future runs but leaves the task in the owner's account and notifies them. »Turn on« resumes execution at the next interval. Manual tasks have no status switch, because they do not run on a schedule.

Since August 2026 the Slack bot, Teams bot and the Microsoft Office add-ins no longer sit scattered across the integration settings but on a dedicated Plugins page. Enabling an add-in or reconnecting a bot happens in one place.

Since August 2026 the agents table in the governance area can also be filtered by model. If you want to check where an expensive or deprecated model is still hard-wired across the workspace, you find the affected agents in one step. The audit log also identifies the acting person more clearly.

Archiving agents and auto-archive (September 2026): Admins archive an agent from the governance area via Archive agent, picking a reason from the list (»No longer used«, »Replaced by another agent«, »Duplicate«, »Seasonal«, »Other«) and optionally adding a note. Restoring happens in the Archived tab via Restore agent. On top of that there is an automatic rule in workspace settings: it is enabled by default in new and existing workspaces, the inactivity window is 1, 3, 6 or 12 months (default 3 months) and the grace period before archival 7, 14 or 30 days (default 7 days). A nightly job checks last usage. Usage means an actual run in chat, Slack, Teams, API, a workflow, an automation or as a subagent; merely opening the agent list does not count. Owners and editors are notified once an agent is scheduled for archival and can prevent it by using the agent during the grace period. The user-side view is in ch. 16.12.

20.5 Governance

Since summer 2026 the admin menu has a dedicated Governance area. It brings together in one place what used to be scattered: reviewing agents, inspecting shared skills and enforcing custom compliance rules.

• Agents view: Every agent in the workspace with its review status, compliance result and risk level. Admins can approve an agent, flag it with a note or disable a problematic version.
• Skills view: All skills shared in the workspace with owner and visibility, plus tabs for summary, sharing, bundled files and the instructions verbatim. Actions such as approve, flag or disable are not available here yet, they have been announced. Until then the route is: review it and talk to the owner.
• Compliance rules: Rules with a name, description, severity (low, medium, high, critical) and a failure condition. Templates for the EU AI Act and GDPR are included, and you can build your own freely. Checks run automatically when an agent is created, on every published new version, and on demand for existing agents. Each rule has its own analytics: executions over time, number of flagged agents, API costs incurred. Rule of thumb on severity: critical only if you would genuinely disable the agent over it.
• Access control: Governance can be opened up to all workspace admins or limited to specific people. Optionally, private agents that were never shared stay out of the compliance view.
• Approval comments: Since August 2026 admins can leave a comment when reviewing an agent. The reasoning behind an approval or rejection stays with the record instead of living in a separate email.

A review queue instead of an endless list: three tabs above the agent list filter down to what actually needs work. Pending review shows agents not yet reviewed, Needs re-review those that changed since their last review, and No owner those without an assigned person, which matters most when someone leaves the company. On top of that you can filter by visibility, integrations, model, and labels.

Integration risk: alongside the compliance result, Governance rates every agent by how risky its available actions are. The Checks tab lists them grouped by integration and then by View, Change, and Delete. Change and Delete count as high risk. An agent with write access to a sensitive system and shared with the whole workspace therefore ranks above a private agent without integrations.

Agent detail view: clicking an agent opens everything a decision needs, without asking its editor for context: Summary (owner, editors, instructions, model, skills, and integrations), Checks, Sharing, Analytics (messages, users, conversations), Costs, and History with all versions and past review decisions.

Three actions, different visibility: approve marks the agent as reviewed. Flag attaches an internal note that editors and users never see, meant as a reminder for the admin team. Disable takes effect immediately and is visible: the owner and editors are notified with your reason, and the agent stays unusable until a new version is published and approved again. Regular users don't see the reason. All three actions are logged in the agent's timeline.

20.6 Default Tool Stack

Since July 2026, admins set a default tool stack in workspace settings under Integrations: Microsoft, Google or »None«. The choice controls which integrations new members see first during onboarding and in recommendations. If your organization runs mostly on M365, this saves the team from hunting between SharePoint and Drive.

• Affects recommendations only. Existing connections and permissions stay untouched if you change the default later.
• Overridable per category. Individual areas (messaging, project management, file storage) can get a different default, for example Microsoft as the suite but Slack for messaging.
• Users win. If someone already has their own connection, it takes precedence over the workspace default.

20.7 Usage Exports

Admin menu → Usage Exports. CSV/JSON export per:

• Model (which model consumes how much)
• Agent (which agent is heavy)
• Project (which projects produce volume)
• User (who are the heavy users)

For large workspaces: the same export is also reachable via API, capped at 1M rows per call. When the selected period gets too large, since August 2026 Langdock suggests splitting the export into monthly chunks instead of failing at the limit.

API costs as CSV: Admins can export API costs over freely chosen date ranges of up to 12 months. Useful for internal cost-center charge-back.

20.8 Recommended Roll-out Order

For workspaces switching to the new fair usage policy or restructuring capacity management:

1. Adjust default model and distribute the three-tier cheat sheet from 3.2 to the team.
2. Activate Extra Usage with a workspace cap of €200 to 300 per month.
3. Keep usage requests on and collect 4 weeks of data.
4. Move 1 to 5 real heavy users to Business Max based on the data.
5. Only then evaluate BYOK if costs still dominate or the use case profile shifts.

What IT security typically cares about. Setup paths sit in the Langdock docs. Here is the option overview.

21.1 SAML / SCIM

21.2 IP Restrictions, Static IP, Session Management

• IP restrictions: Allow workspace login only from specific IP ranges (e.g. only from the company VPN).
• Static IP: Outbound calls from Langdock run via fixed IP addresses. For allowlists in external systems that only accept certain sources.
• Session management: Admins can end all active workspace sessions with one click in Security Settings. Useful during security incidents, when enforcing new auth policies, or on personnel changes. Idle timeouts themselves are configured at the IdP layer (SAML).

21.3 BYOK (Bring Your Own Keys)

Use your own provider API keys instead of the Langdock pool. Available for Anthropic, OpenAI/Azure OpenAI, Google, Mistral.

Authentication: Each deployment can send additional HTTP headers the provider requires under “Custom headers”. For Azure OpenAI, authentication via managed identity was added in August 2026, so the model key no longer has to sit in the workspace as a secret and the permission hangs off the Azure identity instead.

• Own provider limits: You negotiate rate limits and quotas directly with the provider, and Langdock's pool limits don't apply to these calls. Spend limits do apply: in BYOK workspaces Langdock caps usage by monthly amount rather than by messages, at workspace, plan, group, and user level. Details in ch. 20.2.
• Own billing: Model costs run directly between you and the provider, not through Langdock.
• Reduced seat license: since Langdock no longer provides the model.
• Sensible when: Your model consumption permanently exceeds pool limits, or you already have an enterprise deal with the provider.

21.4 Deployment Modes