The MIND MCP server serves public documentation to AI tools. It reads; it never writes. This page sets out what that means for your data and how to use the server safely.
No sign-in, by design
The server is public. It serves only the public MIND docs and the public release sources, and it has no accounts, API keys or OAuth. It never answers 401, and its OAuth discovery paths return 404 so that no client starts a sign-in. If a client offers a sign-in, a bearer token or an API key field, choose none and leave it empty.
Use the official endpoint
The only MIND MCP endpoint is:
Check the URL before you add it, and be wary of look-alike servers and of copies in third-party directories that point somewhere else. The server identifies itself as dev.mindlang/mind in initialize and in its server card.
Read-only by construction
Every tool is marked readOnlyHint: true, destructiveHint: false, idempotentHint: true and openWorldHint: false, and that is what the code does. The server has no write tools and no access to your machine; no tool fetches URLs or contacts other sites, because every answer comes from an index that ships with the server; and it keeps no session between requests. There is nothing to confirm before a tool runs, and calling a tool again has no side effect.
Treat tool output as data
Results come from the public docs, the standard-library sources and the specification. Even so, an agent should treat any text a tool returns as information to read, not as instructions to follow, which is the general rule for content an MCP server returns. Answers should rest on the cited url of each passage, so you can check them against the page.
What is recorded
Each request writes one data point of counts, with every field taken from a fixed list. Request logging for the server is turned off.
| Recorded for each request | Never recorded |
|---|---|
The operation: a tool, prompt or resource kind from a fixed list, or a method name such as tools/list | Query text and tool or prompt arguments |
| The protocol era: 2026-07-28, or earlier | Document ids and resource URIs |
| The outcome, such as ok, tool error, rate-limited or refused | IP addresses and location |
| The search mode (by meaning, from the cache, or by keywords), or the type of an internal error | Client names and versions as sent |
| The client family, from a fixed list of about a dozen common clients, or "other" | Request and response bodies |
| The size of the result and the number of items | Headers, cookies and tokens |
The rate limiter counts requests per IP address (per /64 for IPv6) over one minute so it can enforce the limits; the address is not stored in any log or dataset.
What you put in a tool or prompt argument, such as a search query or code passed to port_to_mind, is used to answer that one request and is not kept. To match by meaning, the server sends the text of a search query to an embedding model on the same edge platform as the server (see Search cache).
Search cache
Search matches by meaning and by keyword. To answer a repeated search quickly, the server keeps the embedding of a search query, a list of numbers that represents its meaning, in an edge cache for up to 7 days. It caches an embedding only after the embedding model returns one. The cache key is a SHA-256 hash of the query; the query text itself is not stored. When the embedding model is unavailable, or one client sends more than 30 searches a minute, search matches by keyword alone (mode: "lexical") instead of failing.
Transport security
Use the https:// URL. Every response carries these headers:
| Header | Value |
|---|---|
Strict-Transport-Security | max-age=63072000; includeSubDomains; preload |
X-Content-Type-Options | nosniff |
Referrer-Policy | no-referrer |
X-Frame-Options | DENY |
Content-Security-Policy | default-src 'none'; frame-ancestors 'none' |
Permissions-Policy | Camera, microphone, screen capture, geolocation, payment, USB and the motion sensors turned off |
Access-Control-Allow-Origin | *, for POST, GET and OPTIONS |
Cross-origin calls are allowed because the content is public and no cookie or credential is ever involved; the Origin rules still refuse plain-HTTP pages on other hosts.
Related pages
- MCP Server: setup for every client
- Limits and Errors: rate limits, error codes and routes
- FAQ: short answers to common questions
- Privacy policy: the mindlang.dev website