The 117 tools in the CognitionDesk MCP, by area: what your agent can do with your account

You connected the CognitionDesk MCP to your agent. Now what can you ask it to do?

claude mcp add cognitiondesk -e COGNITIONDESK_API_KEY=your_key -- npx -y @cognitiondesk/mcp

Update (October 5): @cognitiondesk/mcp 0.10.0 is out, with 122 tools. The five new ones cover YouTube and social: list_social_connections and list_youtube_videos only read; connect_youtube_channel, upload_youtube_video and update_youtube_video write. None of them delete, so the ten destructive tools below are unchanged. The rest of this guide describes 0.9.0.

0.9.0 registers 117 tools, 8 guides and 5 playbooks. It runs with your API key, so your agent can do exactly what that key can do: no more, no less.

Make get_connection_status the first call. It tells you which account the key points to (masked), what plan it's on and which assistants it can manage. Most "nothing works" cases are a missing or revoked key, and this tool says so up front.

The 117, by area

Four areas add up to 46 of the 117: knowledge (13), assistants (12), skills (12) and audience (9).

• Assistants (12). Create, edit, try and publish. try_assistant talks to the assistant and returns what it actually answers. publish_assistant freezes the draft as a version and points the live bot at it; rollback_assistant goes back to an earlier one without touching the draft.
• Knowledge (13). What the assistant knows to answer from: text, URLs, documents, images and GitHub repos. estimate_knowledge_ingest prices an ingest before you run it. search_knowledge searches the same thing the assistant will find.
• Skills (12) and custom tools (5). A reusable behaviour goes in a skill; an action against an external system goes in a custom tool. resolve_skills turns a list of skills into the concrete tools they add and the text they append to the prompt.
• Audience & handoff (9). Who the assistant recognises, through which channels, and which person on the team gets the conversation when it lets go. get_identity_audit shows who proved who they are, who tried and failed, and when.
• Workflows (9), schedules (7) and squads (5). LangGraph graphs with drafts and versions, tasks that run on their own with a run history, and teams of assistants built from a one-line goal.
• Channels & widget (6). create_web_widget gives an assistant an embeddable chat; update_web_widget edits its allowedDomains, the list that decides whether the chat answers on your site.
• The rest: agenda (4), inventory (7), tasks (7), memory & files (3), email campaigns (2), images (1), receivables (1), usage & plans (4), API keys (6) and connection (4). That last one includes use_account, which switches accounts when one server handles several.

Which ones read and which ones delete

53 of the 117 are read-only. They list, search and measure: you can allow them in your client.

10 are destructive, and none of them hides behind a friendly name:

delete_assistant, delete_assistant_contact, delete_assistant_person, delete_handoff_contact, delete_knowledge_entry, delete_skill, delete_custom_tool, delete_task, delete_api_key and revoke_api_key.

Every tool carries those two flags (readOnlyHint and destructiveHint in the protocol), so an MCP client can ask you to confirm only those ten and let the rest through. The other 54 write: they create, edit, publish. Many of them work on a draft that doesn't reach the live bot until you publish it.

Nobody writes the table

The full reference lives at docs.cognitiondesk.com/docs/mcp, and no person maintains it.

The README table used to be written by hand, and it listed about 90 tools when the server already registered 117. Now a script builds the server, connects it to an in-memory MCP client (no network, no API key) and lists what the server actually exposes. Each tool's area comes from the comment above it in the code. The docs page reads that JSON.

And there's a check: if someone adds a tool to the server and doesn't regenerate the docs, npm run tools:al-dia exits 1 and names the difference. The docs can't fall behind the code without something turning red.

Getting started

1. Install it with the command above and call get_connection_status.
2. Ask your agent for list_scopes before guessing a permission: it returns the platform's three permission catalogs in one call.
3. Open the full table at docs.cognitiondesk.com/docs/mcp, with each tool's description exactly as your agent reads it.

If you're missing a tool for what you want to do, write to us through the chat on this page.