Redis Docs MCP

Redis Docs MCP

The redis-docs server exposes the redis.io documentation over MCP as search, fetch, and ask. Those tools carry their own descriptions, and on when to call each one they are accurate — start with search, fetch what you need, keep ask for questions that span several areas. This skill covers what they get wrong or leave out.

When to apply

• You are about to look something up through these tools.

• A result did not look like what the tool description promised.

• A search came back empty, or ask was rate-limited or declined to answer.

Reach for the sibling Redis skills first where they cover the topic — they carry curated guidance that raw docs pages do not. Use these tools to verify specifics and fill gaps.

1. Ids address sections, not pages

The tool descriptions say search returns "pages" and fetch returns "one Redis documentation page". The index is per section, and ids come in four shapes:

What follows from this:

• Repeated urls in one result set are not duplicates. Several hits often share a page url while carrying distinct ids — they are different sections of the same page. Tell them apart by their snippets, not their urls.

• #sec:wN ordinals are index windows. They come back in relevance order, not w0, w1, w2. Don't try to reassemble a page out of snippets; fetch the section you actually need.

• Never construct an id. Valid ids come from a search result or an ask result's sources[].id. Guessing happens to work for commands/<name> and fails on everything with sections.

• Snippets are truncated prefixes of the body, not summaries. When one cuts off mid-word right at the interesting part, fetch on that id is guaranteed to contain the continuation — no need to re-search with different words.

• Cite the page url and name the section in prose. #sec:w1 is an index address, not a web anchor. Ids with a named anchor (#sec:<name>, or #sec:<name>-wN) return a url with a usable fragment; bare #sec:wN ids don't.

Links in returned text point at source files — [`DEL`](/content/commands/del.md) — which are neither web URLs nor fetch ids. Don't cite them or turn them into ids; search for the term instead.

2. metadata.version means "added in", not "current as of"

fetch returns metadata with topic, version, and section. On a commands/* page that version is the release the command was introduced in — commands/setex reports 2.0.0, commands/hexpire reports 7.4.0. Useful for "when did this land", and nothing more: commands/expire reports 1.0.0 while its own body documents a behavior change in 2.1.3 and a Redis 8 interaction with Redis Search. Outside commands/* the field is often not a release at all — a Redis Search page reports stack, a product channel.

So never answer "did this change in Redis N?" from metadata.version. Read it from the body text, which states it explicitly ("Before Redis 8…", "since Redis 7.4"), and cite that.

3. What the corpus covers

The index is redis.io only — but that is broader than it sounds. Per-language client usage guides are indexed and returned (develop/clients/redis-py/connect#sec:…, develop/clients/jedis/connect#sec:…), along with the integrate/… pages, the Redis Software and Redis Cloud operational docs under operate/…, and the Kubernetes operator's generated CRD reference. So search first, including for client-library and Enterprise questions.

Details often sit past the snippet. A search for redis-py's max_connections returns develop/clients/redis-py/async#sec:connection-pools with a snippet that never mentions it, yet the fetched section documents it. For a library's complete signatures, its own API reference or source is the authority.

Genuinely outside the index: GitHub issues and discussions, package-manager pages (PyPI, npm, Maven), and third-party posts.

So fetch the relevant hits before concluding anything is missing, and even then report it as "not found in the results I retrieved" — never as "not on redis.io" or "does not exist".

4. ask costs more than it looks

• Rate-limited per IP at a low per-minute budget, shared by everyone behind the same NAT, so an office or a CI runner can exhaust it collectively. On a rate-limit error, back off before retrying rather than retrying straight away, and ask one narrow question at a time.

• On a refusal, do not retry it reworded. "The documentation does not cover this" is a grounded result, not a parsing failure. Fall back to search + fetch instead.

• ask → sources[].id → fetch when you want the primary source behind a synthesized answer.

Query text from every tool reaches Redis, and from ask the model provider behind the gateway too, so prefer a generic question about the Redis command or behavior over pasting in proprietary code.