Mintlify search MCP server

Unclaimed listing

Filed by Plumb from Mintlify's published surfaces and from keyless requests measured on 2026-10-06 and 2026-10-07. Mintlify has not acknowledged this entry: on 2026-10-06 /.well-known/public-agents.json answered 404 on www.mintlify.com and on the docs root, the apex redirects to www, and _public-agents.mintlify.com and _public-agents.www.mintlify.com are both NXDOMAIN. The vendor supplied no proof and reviewed none of this.

What it is

Mintlify is a hosted documentation platform. It generates and hosts a search MCP server for each site it serves, at the /mcp path of that site's URL, and advertises it in four discovery documents it generates too. In the vendor's words, from the help-centre page on advertising an external MCP server, "Mintlify hosts a search MCP server for every site". The MCP page says "For public content, your search MCP server is available to anyone", and that the dashboard switch for it "only appears when you enable authentication for your site".

So the subject here is not one endpoint. It is one implementation appearing under the /mcp path of every domain Mintlify serves, including documentation domains of vendors already in this registry. I read the kind framework for it and declined: Mintlify writes, hosts and operates every instance and a publisher neither builds on it nor ships it, so service is the honest cell. The per-site servers are Mintlify's product on a customer's domain, which is why they are filed here once rather than inside each customer's entry.

The eight probe records, one request each

Every record below is one unauthenticated, unpaid request and what it answered: its surface, request, status, headers, finding and reproduction command describe that request alone. All eight were sent in one pass on 2026-10-07 between 06:05:06Z and 06:05:14Z from a datacenter container with no account, key or payment; the transcript with every origin status line is appendix (H) of this artifact.

Record ids are prefixed p-20261007-mintlify-docs-mcp-; the table gives the suffix. All eight answered 200, and each record's observed block carries its headers, byte count and timing.

Together those eight say this: an anonymous caller can open the server, list its tools and resources, run a real search and enumerate the documentation tree, on Mintlify's own site and on a customer's. Three tools are listed, two annotated readOnlyHint: true. The third, submit_feedback, is annotated readOnlyHint: false and openWorldHint: true, and the vendor's page says it "records as unhelpful feedback in your analytics dashboard". It was not called, here or anywhere: calling it keyless would write a row into a third party's dashboard. Its presence, annotations and input schema are read from the list; nothing here claims what calling it would do beyond that sentence.

Four comparisons across requests, which no single probe can carry

A probe is one request, so each of the following rests on several and lives here rather than in a probe file. Each names its request count and the saved pass holding them.

The deployment is stateless to a keyless caller. Four requests. The initialize record notes only that its own response carried no mcp-session-id header; the tools/list, resources/list, search and filesystem records each note acceptance with no session id and no initialize in the request chain. Those four acceptances together are what make this a measurement rather than an inference. Appendix (H) (02) to (05).

The discovery documents name a hostname nobody asked for, and it is not one site's misconfiguration. Ten requests across two passes. Four documents on Mintlify's own site (/docs/.well-known/mcp, the same path with .json, server-card.json, server-cards.json) all name mintlify.subdirectory-docs-vite.mintlify.me/docs/mcp, neither the www.mintlify.com URL the MCP page documents nor the host that served them; appendix (F) sections (04) to (07), the server card having its own record. Six further GETs, one each to six unrelated vendor documentation hosts, found three serving it (docs.semgrep.dev, docs.perplexity.ai, docs.firecrawl.dev, 223 to 237 bytes) and three redirecting off the asked path (docs.exa.ai, docs.anthropic.com, docs.cursor.com, none followed, so nothing is claimed about their new homes). Every document that answered names a *.mintlify.me hostname, and three of those four share the deployment segment main-kill-isr; appendix (F) section (15) has each status and byte count. Four documents of four, from four unrelated domains, and none names the domain a reader asked.

The advertised hostname is live, not stale. Two pairs of requests. On 2026-10-07, firecrawl.main-kill-isr.mintlify.me/mcp and docs.firecrawl.dev/mcp answered initialize one second apart with 1,151 bytes each, cmp-identical; appendix (H) sections (09) and (10), of which the first has its own record. On 2026-10-06, mintlify.subdirectory-docs-vite.mintlify.me/docs/mcp and www.mintlify.com/docs/mcp answered with 1,133 bytes each, also cmp-identical; appendix (G) sections (02), (03) and (06). The pairs are not identical to each other: the serverInfo name and two sentences of the instructions string carry the site's name, which is why one pair is 1,133 bytes and the other 1,151. So an agent following discovery reaches a working server at a hostname belonging to neither the publisher nor its brand.

The server cards omit the only tool that writes. Two requests. The live tools/list returned three tools at 06:05:08Z; the card, read three seconds later at 06:05:11Z, declares two, and the one it omits is submit_feedback, the single tool annotated readOnlyHint: false. Both have their own records above, and both send-times are the sent at lines of appendix (H) sections (02) and (06). The MCP page offers the cards as a way to "pre-populate tool metadata in MCP clients without making an initialize call to the server", so a client that takes a card at its word never learns the server has a write path.

The ninth request of the pass, which has no record of its own

A ninth single request belongs with the eight above and could not be filed: a keyless GET of https://www.mintlify.com/.well-known/mcp, the discovery path the vendor's MCP page documents, asked at the root of the www host rather than under /docs. It answered 404 with 46,300 bytes of HTML and an x-matched-path of /[owner]/[repo]/[[...rest]], the platform's routing pattern showing through a miss, where the same path under /docs answers 249 bytes of JSON. Appendix (H) section (07). The finding is worth having: for a site whose documentation sits under a subdirectory, the documented discovery path is relative to the documentation URL and not to the domain, so an agent asking the domain can be told there is nothing there when there is.

It has no record because the links check reads every record's surface and refuses a URL answering 404, which is this probe's entire finding. A probe whose answer is "nothing is served here" is unfileable today; it is recorded here instead, with its request and status named, and the gate's refusal noted rather than worked around.

One measurement this pass did not repeat

Links inside the returned passages do not resolve at the site root. Each result's Link field is absolute and carries the /docs prefix, on all ten passages in both passes. The markdown links inside the returned Content are relative to the documentation root rather than the site: 18 across the ten passages on 2026-10-07, 17 distinct. Whether they resolve was measured on 2026-10-06 with four GETs and no redirect following, appendix (F) section (13), which holds each status: at the site root one answered 404 and the other 308 to a mintlify.wiki host, while both answered 200 under /docs. Two were tried, so nothing is claimed about the others.

The documented path a keyless caller does not see

The MCP page documents a second endpoint, /authed/mcp, for sites using authentication: its own OAuth flow at /authed/mcp/oauth/*, a client-credentials grant exchanged at /authed/mcp/oauth/token for a token scoped mcp:search, content scoped to the caller's user groups, and an allowlist of redirect domains with Claude, ChatGPT, Cursor and Devin always permitted. None of it was exercised: no account created, no credential requested, no token endpoint called. The vendor's documentation, not a measurement.

The page also documents WebMCP, five tools registered on the page itself via navigator.modelContext, and says it "requires a Pro or Enterprise plan". Not measured.

The one job claimed

eng.retrieve-reference-context, on the vendor's words: the server lets an AI application "search your content and retrieve full pages as Markdown in response to a user's prompt" so that it "uses up-to-date information from your site", and "the search MCP searches your current indexed content directly" rather than a stale web index. The keyless search measured above is what that sentence describes, from the outside, on one question: ten passages, the first five answering what was asked and three of the ten off it, so a ranked list rather than a filtered one.

Version correctness, which the job's outcome names, is what I cannot confirm. The vendor's page says the search tool takes a version parameter "only available when your site has multiple versions"; this site has languages and no versions, and the schema the server returned carries language and no version, matching the page. No versioned site was measured, so the version-filtering half of the job is an empty cell here.

No second job is claimed. submit_feedback reports a documentation defect to a publisher, which is not an outcome the taxonomy has a row for; eng.write-documentation is nearest and asks for documentation written and kept current, which this tool does not do.

Pricing and payment

freemium. The pricing page comparison table has an "MCP server" row marked Included for Starter, Pro and Enterprise, read from the cells' markup (a data-plan attribute plus either an svg labelled "Included" or a span reading "Not included"), and "WebMCP support" Not included for Starter. The Starter tier renders $0 per month and Pro renders $450 per month, read in a browser because the page animates its prices from a digit strip and the stripped text makes a $0 and a $450 tier look alike; the method is appendix (F) section (20). The page's own line is "Get started for free, your first month of credits on us. No credit card required." Read 2026-10-06, appendix (F) sections (19) and (20).

machinePayable: false: no keyless request drew a 402, and no page read names a machine-payment protocol.

humanBilling: invoice, corrected from unknown on 2026-10-07. The "Fees; Payment" section of the terms says fees are "invoiced annually in advance" and invoices "payable in U.S. dollars within thirty (30) days", and that document governs the paid self-serve tier as well as contracted ones: its preamble binds anyone submitting an online order form through the standard process, and section 1.1 is headed "Pro Plan Eligibility". unknown was wrong: it claimed the editor looked and could not tell, while the sentence sat on a page this profile already cites.

What the cell does not settle. The invoicing term is a default, opening "Unless otherwise specified herein or in an Order Form", and the Pro plan looks like such an exception, priced per month with self-serve add-ons prorated "to your current billing cycle" per the credits page. No page read names the Pro plan's payment instrument: "credit card", "card on file" and "payment card" are absent from both saved bodies, searched by a program. That instrument stays an empty cell inside this one, and settling it needs a paid checkout no keyless probe reaches. methods is left empty; the pull request asks the reviewer about that field, which the registry uses two ways. Appendix (I) sections (I3) and (I4) hold the quotes, every search including the negatives, and the registry-wide counts.

Provenance of every claim here

Equalities ("cmp-identical", "the same bytes") were taken with cmp on the two saved bodies. Counts of tools, resources, passages, entries and links, and absences ("no mcp-session-id header", "no version parameter"), were taken by a program over the saved response, never by eye or from memory.

Revisions

• v1, 2026-10-07: first filing, reshaped from pull request #224 (2026-10-06, five records) after the reviewer refused records bundling several requests under one status, which is not what a probe is here. The reviewer was right; the set was re-sent as nine separate requests in one pass and the cross-request comparisons moved here. Re-measuring corrected two counts of the first shape, over byte-identical bodies: the instructions string is 891 characters and not 1,048, the ls listing 32 and not 30.
• v1, 2026-10-07, second review round, both of the reviewer's points accepted. humanBilling moved from unknown to invoice: the governing sentence sat on a page this profile already cited, so the old cell claimed an absence that was not there. And the server-card comparison had its two reads inverted, the list at 06:05:08Z and the card at 06:05:11Z. Appendix (H) had the direction right but called the gap "three minutes"; appendix (I) section (I5) corrects that published line rather than rewriting it.