Ctrlk
For the complete documentation index, see llms.txt. This page is also available as Markdown.

MCP Tools

Audience: Developers and integration engineers

Purpose: Look up each MCP tool's inputs, outputs, and behavior.

Overview

The Kizen MCP server exposes three tools for working with Knowledge Catalogs, which are sets of information sources that you can share across your business. Admins control which roles and team members can query each catalog. A connected assistant discovers these tools automatically and calls them as needed to answer questions from the catalogs you are allowed to access.

Why Would I Use These Tools?

Use these tools to let an AI assistant answer questions from your Kizen knowledge without anyone copying content out of Kizen. The assistant finds the right catalog, searches it, and pulls the source document, all scoped to what you are allowed to see.

listCatalogs

Lists the Knowledge Catalogs the connected user can access.

Inputs: none.

Returns: a list of the catalogs the user can access, with a total count, and for each catalog its id, api_name, and description.

searchCatalog

Runs a semantic search across one or all accessible catalogs and returns the most relevant content.

Inputs:

  • query (string, required): the search text.

  • n (integer, optional, default 10): maximum number of results.

  • catalog (string, optional): limit the search to one or more catalogs, by API name or UUID. Accepts multiple comma-separated values. To search all accessible catalogs, omit the catalog input.

Returns: matching content chunks as plain text, each with a source description, a snippet, and a relevance score.

getDocument

Retrieves the full text of a single knowledge source.

Inputs:

  • source_id (string, required): the source identifier, obtained from a searchCatalog result.

Returns: the source's metadata and its full, concatenated content.

Behavior and Details

  • Response format: All tools return plain text, not JSON, optimized for AI consumption.

  • Tool annotations: Each tool advertises MCP annotations such as readOnlyHint: true and destructiveHint: false, so clients can tell these are safe, read-only tools.

  • Permissions and scope: Results include only catalogs the user has View permission on, scoped to the authenticated business.

  • Performance: A search adds roughly 200 to 500 ms for embedding.

  • Errors: An unauthenticated request returns a JSON-RPC error. An unknown configuration slug returns 403.

What's Next

For how the server and authentication work, see Build with MCP, or follow Connect an AI Assistant to connect a client.

Related Topics
PreviousConnect an AI Assistant
NextPermissions

Last updated

Was this helpful?