MCP Integration Guide for Claude & Cursor

What is iknow MCP?

iknow MCP is a remote MCP server for working with agent knowledge on iknow.dev directly from an AI assistant. In addition to keyword and semantic search, owners can create, edit, and delete knowledge. Access control matches the web app: each agent's owner / member / public roles apply.

Connecting from Claude

Connect as a custom connector in Claude (web, desktop, or mobile).

1. Sign in to iknow.dev and open the "Agent integrations" button in the header of the chat screen (or the knowledge editor).
2. Copy the remote MCP server URL under "Install to Claude".
3. In Claude, go to Settings → Connectors and choose "Add custom connector". Enter the copied remote MCP server URL and add the connector.
4. Click "Connect" on the connector. The iknow.dev authorization screen opens; review and approve it.
5. Enable the connector in a chat, and Claude can use the knowledge of every agent you have access to.

The OAuth client is registered automatically on connection, with Claude's callback URLs (both claude.ai and claude.com) allowed.

Connecting from Cursor

Click "Install to Cursor" in the same "Agent integrations" modal. Cursor opens and registers the MCP server in one click (credentials are configured automatically).

Available tools

Read-only tools are available to owners, members, and anonymous users of public agents; write tools require the owner role (adding comments is also open to members). The one exception is create-agent, which any logged-in user can call regardless of their role on existing agents. Every tool declares the standard MCP annotations (readOnlyHint for read-only tools, destructiveHint for write tools).

search-knowledge

Searches knowledge by keyword, semantic (vector) similarity, or both. Omit query to list an agent's knowledge in numbering order (filterable by tag and hierarchy depth).

get-knowledge

Retrieves knowledge content, related knowledge links, and the heading outline. Pass include=["comments"] to also return the section comments (filterable with comments_source="health_check" for duplicate health-check review).

get-attachment

Retrieves an image or file attached to knowledge content. Just pass the URL from the body: images of 3MB or less come back as the image itself, so the AI can see it right away. Larger images and files such as PDF or Word return a signed download URL (valid for 15 minutes) to fetch them from.

get-reference

Returns the list of accessible agents plus detailed tool and Markdown guides.

create-agent

Creates a new agent (code and name required; description and prompt optional). The creator becomes the sole owner and the agent is created private. Available to any logged-in user.

create-knowledge

Creates a new knowledge entry.

append-knowledge

Appends content to the end of existing knowledge.

create-attachment-upload-url

Issues a signed upload URL (valid for 15 minutes) for attaching images or files. Multipart POST to that URL, then pass the response url to append-knowledge attachments (for images totaling 3MB or less, prefer append-knowledge with inline base64 data).

update-knowledge

Updates knowledge content (full replace, anchor replace, section replace, or batch edits). It can also update the title, tags, and number and link/unlink related knowledge — including metadata-only updates that leave the content untouched.

delete-knowledge

Deletes a knowledge entry (soft delete).

add-comment

Adds a comment to a knowledge section (available to members as well).

Example prompts

Once connected, you can ask Claude things like:

• "Search the knowledge of (agent name) for information about 'pricing' and summarize the key points" — searches knowledge, reads the content, and summarizes it.
• "Register today's meeting decisions as new knowledge with the tag 'minutes'" — creates knowledge from the conversation (owner only).
• "Append today's decisions to the meeting-minutes knowledge of (agent name)" — appends to the relevant section of existing knowledge (owner only).

Troubleshooting

Connection (OAuth authorization) fails

An incorrect Client ID or client secret is the most common cause. The secret is shown only once at issuance; if you no longer have it, delete the old client in "Agent integrations" and issue a new one.

Tools are missing, or write operations fail

Available tools depend on your role for each agent. Write tools require the owner role (adding comments is open to members).

A "rate limit exceeded" error is returned

Making many tool calls in a short period triggers a temporary limit. Wait a while and try again.

The agent is reported as not found (404)

Private agents are accessible only to their owners and members. Check the spelling of the agent code and your membership of the agent.

Support

If the issue persists, please contact us at the address below.
contact@iknow.dev