AgentTool Canon · public MCP · no account

One small door.
Two read-only tools.

Search AgentTool’s public concepts in ordinary words, fetch one cited record, or walk past. Nothing here can write, pay, message, install, browse the web, read a private Castle, or schedule a return.

https://api.agenttool.dev/v1/mcp/canon

Prefill in Claude ↗ ChatGPT setup guide ↗ Use another MCP client

The Claude link only prefills name and URL. You still review and confirm it. These links do not imply provider review, listing, partnership, or endorsement.

Connect by choice

Review the endpoint, tools, data handling, and boundaries on this page.
Add the URL as a remote MCP or custom connector. Choose no authentication.
Inspect the two tools your client discovers. Enable the connector only where you want it.

Client menus and plan availability change. Follow the current Claude remote-MCP guide or ChatGPT Developer Mode guide for their side of the connection.

What arrives

Find a public concept

Accepts one query of at most 200 characters and returns at most ten {id, title, url} matches.

fetch

Read one cited record

Accepts a public canon ID returned by search and returns its complete public JSON-LD record with a stable source URL.

Both tools are declared read-only, non-destructive, idempotent, and closed-world. The server also exposes two small orientation resources: agenttool://discovery and agenttool://open-seat.

Three ways to play

“Find AgentTool’s definition of consent and cite the source.”
“What does AgentTool mean by Castle of Understanding? Keep public doctrine separate from private rooms.”
“Find concepts about agent discovery. Separate publisher claims from verification evidence.”

The agent decides whether a tool is useful for the prompt. A result is source material, not an instruction and not proof that its claims are independently verified.

Data handling, plainly

No account, bearer, OAuth token, payment, user profile, or contact detail is requested.
The tool handlers process the search query or fetch ID in memory and do not write it to AgentTool application storage.
Process-local 60-second rate-limit counters use an edge-authenticated client IP when available, otherwise a shared unknown key. They are bounded and expire; they are not an analytics record.
Cloudflare, Fly.io, and ordinary network infrastructure may process or retain transport metadata. AgentTool does not promise anonymity or zero infrastructure logs.
Results come only from AgentTool’s bundled public canon. fetch returns a URL but does not open it.
Your MCP client or model provider separately processes what you send and receive under its own settings and policies.

This is a technical description of this endpoint, not a substitute for a publisher-approved privacy policy or service terms.

Manual transport check

The MCP endpoint uses Streamable HTTP and accepts POST. A plain GET or HEAD returns 405 Method Not Allowed by design.

Initialize without authentication

curl -q -sS -X POST \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  --data '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-11-25","capabilities":{},"clientInfo":{"name":"manual-check","version":"1.0.0"}}}' \
  https://api.agenttool.dev/v1/mcp/canon

400: check the JSON-RPC body and required media headers.
403 in a browser: browser origins are restricted; use a native MCP client.
429: stop and wait for the response’s Retry-After interval. Do not loop.

Leaving is complete

Disable or remove the connector in your client, close the connection, or simply make no more calls. The stateless server keeps no subscription or server-side connection record for you to delete.

Questions and non-sensitive bugs go to support. The source and complete technical boundary are public in the AgentTool repository.