Adding Connect to ChatGPT
Point a ChatGPT connector at https://connectbyjbrh.com/mcp with no authentication. Connect publishes a standard remote MCP server and nothing vendor-specific; whether your account and client can add a custom connector is decided by them, not by Connect, and this page will not guess at their current menus.
What Connect provides#
| Requirement | Connect |
|---|---|
| A single HTTPS endpoint | https://connectbyjbrh.com/mcp |
| Streamable HTTP transport | Yes, POST-only, revision 2026-07-28 |
| Older revisions accepted | 2025-11-25, 2025-06-18, 2025-03-26 |
| Authentication | None required and none supported |
| Tool list | Ten read-only documentation tools |
Tools named exactly search and fetch | No — they are search_public_docs and fetch_public_doc, and nothing aliases them |
| Write or action tools | None |
| Server-sent events / streaming | No; the revision removed the standalone stream |
That last-but-two row matters if you are wiring Connect into a mode that expects a server to expose two particular tool names. The names here are the ten returned by tools/list, and there is no alias layer, so a mode that requires other names will report the server as unsuitable rather than working partially.
Setting it up#
In the connector settings of your ChatGPT client, choose to add an MCP server and paste
https://connectbyjbrh.com/mcp.Result No credential is requested by the server. If the client offers authentication choices, pick none.
Save and let the client enumerate the tools.
Result Ten tools appear. Each is annotated
readOnlyHint: true, which a client may use to skip a confirmation prompt.Ask something the documentation answers, and watch which tool runs.
Result
search_public_docsis the usual first call; a follow-upfetch_public_docreads the winning page in full, which is where the actual answer comes from.
Limits worth knowing before you build on it#
- 120 requests per minute per address, answered with HTTP
429. NoRetry-Afterheader is sent, so choose your own backoff. - Request bodies over 256 KiB are refused with
413. - Search returns at most 25 results however large a
limitis asked for, and there is no cursor to page with. - A fetched page body is capped at 60,000 characters.
- Batched JSON-RPC messages are refused; send one per request.
None of these are negotiable per client, and none of them change with an account, because there is no account. If a use case needs more than that, the same corpus is available as static files — the machine-readable documentation lists them — and fetching one manifest is cheaper than a hundred tool calls.
Alternatives if the connector route is closed to you#
Connector availability varies by plan and by client version, and Connect has no influence over it. When adding one is not possible, two routes reach the same information. The public API answers the same questions over ordinary GET requests, which any environment that can make an HTTP call can use. And llms.txt on this site plus the Markdown mirrors give a model the whole map without any tooling at all.
Whichever route you take, the content is identical: all of them call the same search and fetch functions over the same generated corpus, so an answer cannot depend on which door it came through.
Questions#
Does Connect have an official ChatGPT app or plugin?
What Connect publishes is a standard MCP endpoint, an A2A agent and a public HTTP API. There is no vendor-specific package beyond those, and this page describes the endpoint rather than any store listing.
Can a ChatGPT connector act on my Connect workspace?
No. Every published tool is read-only and documentation-only, on every client. There is no authenticated tool for any connector to reach.
Why does my client say the server does not support the required tools?
Some client modes require particular tool names. Connect's are search_public_docs and fetch_public_doc; nothing here renames them. Use the server as a general connector, or call the public API directly.