Public MCP tools

Q: How do I find what Connect cannot do?

Call list_capabilities with status set to foundation or not_yet, or call get_product_status for the whole manifest. This is the most useful thing on the server: it is the difference between an agent that says Connect records calls and one that says the model exists but the live carrier does not enable it.

Q: Why does search_public_docs return at most 25 results?

Because the index caps a result set at 25 however large a limit you ask for. Search is ranked rather than paged — there is no cursor and no next link. Narrow the query, or use list_capabilities to walk a kind in order.

Ten tools, all read-only, all reading the generated documentation corpus: search_public_docs, fetch_public_doc, list_capabilities, get_capability, get_workflow, search_glossary, get_technology, get_protocol, get_product_status and get_changelog. Each result carries both a text block and structuredContent, so a model reads prose and a program reads JSON.

Status: Available What this means
Audience: developer
Last verified: 2026-09-10
Product version: 6.3.2

The tools#

Tool	Required	Optional	Returns
`search_public_docs`	`query`	`limit` 1–25 (default 8), `kind`, `status`, `channel`	Ranked page summaries: id, url, title, description, summary, status, audiences, channels, updated, score
`fetch_public_doc`	`id_or_url`	`body` (default `true`)	One page, with its Markdown body and its `related` list
`list_capabilities`	—	`status`, `channel`, `audience`, `limit` 1–400 (default 200)	Capability pages in path order
`get_capability`	`id`	—	One capability page in full
`get_workflow`	—	`id`	One workflow page, or every workflow page when no id is given
`search_glossary`	—	`query`, `limit` 1–25 (default 8)	Matching glossary entries, or every defined term when no query is given
`get_technology`	—	`id`	One technology page, or the list of them
`get_protocol`	—	`id`	One protocol page, or the list plus the version of each protocol Connect implements
`get_product_status`	—	—	The status manifest: the five status words, every capability with its status, and the corpus counts
`get_changelog`	—	—	The published changelog of user-visible changes

The kind filter accepts capability, workflow, technology, protocol, troubleshooting, guide, glossary, comparison, research, developer and hub. The status filter accepts the five status words. The channel filter accepts email, phone, whatsapp, sms and softphone. An empty string means no filter, which is why every enum includes one.

What a result looks like#

{"jsonrpc": "2.0", "id": 7, "result": {
  "content": [{"type": "text", "text": "{ \"query\": \"held draft\", … }"}],
  "structuredContent": {
    "query": "held draft",
    "results": [
      {"id": "C0142", "kind": "capability",
       "url": "https://connectbyjbrh.com/docs/email/held-drafts/",
       "title": "Held drafts", "status": "available",
       "audiences": ["both"], "channels": ["email"],
       "updated": "2026-09-10", "score": 12.418}
    ]}}}

Both forms are always sent for a successful call. The text block exists so a model can read the answer without a parsing step; structuredContent exists so a program does not have to parse prose back into fields. They are generated from the same object, so they cannot disagree.

A tool that cannot answer returns a normal JSON-RPC *result* with isError: true and a sentence saying what to do instead — a missing query, or a reference that matches no page. That is a tool-level failure, not a protocol failure, and it is deliberately not a JSON-RPC error object: the model is meant to read it and try again. Protocol failures are on MCP error shapes.

Fetching by id or by path#

fetch_public_doc takes a page id such as C0142, a site-relative path such as /docs/phone/inbound-calls/, or a full canonical URL. A missing trailing slash, a query string, a fragment and a trailing index.md are all normalised away, so the reference you copied out of a search result works whatever form it was in.

Ids are stable and opaque. Do not parse one: the letters and digits are an identifier, not a classification, and a page's kind is a field you can read directly. The body returned is the Markdown twin of the HTML page, generated from the same source, capped at 60,000 characters and carrying no navigation chrome for a model to mistake for content.

Resources and prompts#

Four resources are advertised, each a document already served from the website: https://connectbyjbrh.com/llms.txt (a map of the documentation), /docs-manifest.json (every public page as JSON), /docs-data/status.json (the status of every documented capability) and /.well-known/agent-card.json (the A2A card). resources/read returns them with their real content types; resources/templates/list returns an empty list, because there are no templated resources.

Two prompts are offered. explain_connect takes a question and asks the model to answer only from the documentation, cite the page URL, and say plainly when the status is foundation or not_yet rather than describing it as available. check_capability takes a capability and asks the model to check the claim against get_product_status before repeating it.

Questions#

How do I find what Connect cannot do?

Call list_capabilities with status set to foundation or not_yet, or call get_product_status for the whole manifest. This is the most useful thing on the server: it is the difference between an agent that says Connect records calls and one that says the model exists but the live carrier does not enable it.

Why does `search_public_docs` return at most 25 results?

Because the index caps a result set at 25 however large a limit you ask for. Search is ranked rather than paged — there is no cursor and no next link. Narrow the query, or use list_capabilities to walk a kind in order.

Can a tool call change anything in a workspace?

No. There is no write tool, no workspace scope, and no credential that would grant one. See Workspace MCP tools for what that means in practice.