Using Connect from any MCP client

One call, end to end#

curl -sS https://connectbyjbrh.com/mcp \
  -H 'Content-Type: application/json' \
  -H 'MCP-Protocol-Version: 2026-07-28' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"search_public_docs",
                 "arguments":{"query":"call recording","limit":3},
                 "_meta":{"io.modelcontextprotocol/protocolVersion":"2026-07-28"}}}'

No Origin header is sent, which is correct for anything that is not a browser. No cookie, token or session identifier is involved. The reply is one JSON object with a result containing both a text block and structuredContent.

The methods a client may send#

Anything else raises method-not-found. The initialize result is worth reading even if you skip the handshake: its instructions field is the text that tells a model several Connect capabilities exist as foundation rather than as usable features, which is the single most useful correction a client can pass through to a model.

Writing a minimal client#

1. Decide whether to handshake. If your client is new, skip initialize and include _meta on every request. If it is an existing client for an older revision, send initialize and use the version in the reply.

   Result Either path is accepted; mixing them is fine too.

2. Send optional headers only if you mirror them exactly. MCP-Protocol-Version, Mcp-Method and Mcp-Name must match the body, or the request is rejected with -32020.

   Result Sending none of the three is the simplest correct choice.

3. Parse the body before the status line. A response can be HTTP 200 and still contain an error object — an unknown tool name arrives that way.

   Result Your client records failures accurately instead of counting a call to a non-existent tool as a success.

4. Handle 429 by waiting. The limit is 120 requests per minute per address and no Retry-After is sent.

   Result A widening delay clears it; an immediate retry does not.

5. Cache what you fetch, keyed on the page's updated date.

   Result The corpus changes on a deploy, not per request, so a cached page stays correct until the date moves.

Reading a resource#

{"jsonrpc": "2.0", "id": 2, "method": "resources/read",
 "params": {"uri": "https://connectbyjbrh.com/docs-data/status.json"}}

The four resource URIs are absolute and are the same documents the website serves, so a client with an HTTP fetcher can read them directly instead — llms.txt for the map, docs-manifest.json for every page as JSON, docs-data/status.json for capability status, and the A2A card. Going through resources/read buys you the media type and one code path for everything; going direct buys you edge caching.

Questions#

Related#

• The Connect MCP server
• Public MCP tools
• MCP error shapes
• Designing a resilient client
• JSON-RPC 2.0
• Streamable HTTP