Redis launches a public, keyless docs MCP server for AI agents

Redis has built Redis Docs MCP, a public MCP server at redis.io/mcp that gives coding agents three tools for reading the Redis documentation without authentication. According to Redis, the aim is to stop agents from answering from stale training data or scraped HTML.
- Redis launched Redis Docs MCP, a public MCP server at redis.io/mcp with no authentication.
- It offers three tools: search (up to 8 pages with stable ids), fetch (full page text and metadata) and ask (prose answer with sources).
- search and fetch query the Redis index directly and keep working if the model or gateway is down; ask runs through Context Retriever and a language model.
- Per-caller request text is not logged and telemetry is aggregate-only, according to Redis.
- Four engineers built it between April and July 2026, starting from a three-week demo.
The problem: confident answers from outdated sources
Redis engineer Vishal Bala describes the issue in a company blog post. Ask a coding agent how to configure eviction in Redis 8 and the answer often does not come from authoritative docs. The agent either recalls something from pretraining or scrapes rendered HTML from redis.io at whatever freshness, including marketing copy and old version notes.
That matters because, according to Redis, Redis 8 changed several defaults. A model trained before the release can give the Redis 7 answer with full confidence. Redis says the existing options fall short: web search returns pages without stable identifiers, so an agent cannot cite what it read or return to it, and client-side retrieval forces every framework to reimplement chunking, embedding and freshness against a corpus it does not own.
Three tools, no sign-up
Redis Docs MCP serves three tools over the Redis docs, unauthenticated, to any MCP client. Four people built it over one quarter, from April to July 2026, according to the post.
- search(query) returns up to 8 matching pages, each with a snippet, a title, a URL and a stable id. It is the default for any Redis question.
- fetch(id) returns the full text plus topic, version and section metadata for one page, for when a snippet is not enough or the agent wants to quote exactly.
- ask(question) returns a prose answer plus the source documents behind it, aimed at questions that span several areas where no single page answers.
Redis kept the count at three on purpose. Every tool description is folded into the client's context for the whole connection, so a fourth tool would cost every caller tokens on every call. The company also stresses that this is not the official Redis MCP Server, which is self-hosted and lets an agent run commands against your own Redis instance. Redis Docs MCP is hosted by Redis, read-only, and touches only its own docs index.
Why keyless, and what follows from it
Redis describes the lack of authentication as a deliberate choice. The docs serve both the open-source community and Enterprise customers, and the company says a signup in front of a docs lookup would shut out most of the first group. Other providers made different calls: Microsoft Learn, Cloudflare and Upstash's Context7 run docs MCP servers, while Supabase authenticates every caller, which Redis calls right for a server acting under a developer's permissions.
Because Redis cannot tell who is calling, per-caller request text never reaches logs and telemetry is aggregate-only. Neither internal ranking scores nor non-fetchable identifiers cross the wire. Since all three tools arrive as one JSON-RPC POST to /mcp, per-tool rate limits have to live in the application, as the edge cannot tell search from ask without parsing the body.
Two paths to one corpus
search and fetch follow the shape OpenAI established for docs MCP servers, down to naming the snippet field text instead of content. Redis says ChatGPT Deep Research connectors require that contract, so those clients work with no integration effort on either side. Both tools query the Redis index directly through RedisVL and issue their own FT.HYBRID query against the Redis Query Engine. Redis says this lets it tune the text field, the scorer, the fusion method and the candidate count, and that it swept all four.
ask takes a different route. It is the only tool that touches Context Retriever and a language model, running the same retrieval-and-synthesis pipeline as the earlier marketing demo, shared as a library rather than called over HTTP. A model holds Context Retriever's generated tools and decides which to call, so the model drives retrieval. The payoff, according to Redis, is availability: search and fetch keep serving when the gateway or the model is unavailable. The cost is that the two paths must be debugged separately and can disagree about the best match for a query.
Behind both sits an index with one writer and two readers. The ingest pipeline reads Hugo markdown from the redis/docs repository, strips front matter and shortcodes, chunks long pages by heading, embeds each chunk and imports the result through Context Retriever. Neither the demo chat endpoint nor the MCP server owns the index lifecycle, the schema or embedding generation.
From demo to reusable interface
The project began as a three-week effort in late April 2026 to put a grounded docs assistant on the product page for Context Retriever. It used a FastAPI service streaming server-sent events and roughly a hundred curated docs pages, and went live at the start of June 2026. Redis says that demo proved the pipeline but said nothing about the interface. The quarterly plan named the risk to avoid as "spending the quarter polishing the web demo without producing a reusable agent-facing interface."
For developers building agents, the practical upshot is a single MCP endpoint that returns stable ids they can cite and fetch again. The post also covers design review, a worked example of making ask return citations the model cannot fabricate, and retrieval-quality measurement, and ends with instructions for connecting a client.
Redis: How we built the Redis Docs MCP for agents

