redis-docs-mcp — скилл для Claude Code: функции и установка
Скилл redis-docs-mcp для Claude Code объясняет, как читать результаты поиска по документации Redis через MCP. Он включается перед поиском или когда результат search, fetch либо ask вызывает вопросы.
MIT 101 установок ★ 178 обновлён 2026-09-08
npx skills add redis/agent-skills --skill redis-docs-mcp -a claude-codeЧто делает redis-docs-mcp
Скилл разбирает, как устроен индекс документации Redis и что возвращают инструменты search, fetch и ask. Он объясняет, почему несколько результатов могут вести на один URL, а короткий фрагмент — быть частью раздела, и как использовать ID результата для fetch.
Также он помогает толковать metadata.version и искать сведения о клиентах, Redis Software и Redis Cloud. Индекс включает материалы с redis.io, но не GitHub issues, страницы пакетных менеджеров и сторонние публикации. Для работы нужны доступные инструменты MCP-сервера Redis Docs.
Какие задачи решает
- найти нужный раздел документации Redis по результатам поиска
- разобраться, почему несколько результатов ведут на один URL
- получить продолжение короткого фрагмента через
fetch - не принять
metadata.versionза текущую версию Redis - понять, какие материалы входят в индекс Redis Docs
Примеры запросов агенту
- Найди в документации Redis сведения о сроке появления команды
HEXPIRE. - Результаты
searchведут на один URL. Объясни, как выбрать нужный раздел. fetchвернул короткий фрагмент. Получи полный текст нужного раздела.
Как установить redis-docs-mcp
1Через Skills CLI (Vercel)
npx skills add redis/agent-skills --skill redis-docs-mcp -a claude-code2Вручную: положить файлы
~/.claude/skills/redis-docs-mcp/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/redis-docs-mcp/
1Через Skills CLI (Vercel)
npx skills add redis/agent-skills --skill redis-docs-mcp -a codex2Вручную: положить файлы
.agents/skills/redis-docs-mcp/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add redis/agent-skills --skill redis-docs-mcp -a cursor2Вручную: положить файлы
.cursor/skills/redis-docs-mcp/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install redis/agent-skills skills/redis-docs-mcp2Вручную: положить файлы
.github/skills/redis-docs-mcp/SKILL.md1Через Skills CLI (Vercel)
npx skills add redis/agent-skills --skill redis-docs-mcp -a gemini-cli2Командой платформы
gemini skills install https://github.com/redis/agent-skills --path skills/redis-docs-mcpВручную: положить файлы
.devin/skills/redis-docs-mcp/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Через Skills CLI (Vercel)
npx skills add redis/agent-skills --skill redis-docs-mcp -a opencodeClaude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о redis-docs-mcp
Как установить redis-docs-mcp в Claude Code?
Выполни npx skills add redis/agent-skills --skill redis-docs-mcp -a claude-code. Файл скилла находится по пути ~/.claude/skills/redis-docs-mcp/SKILL.md.
Что делает redis-docs-mcp?
Объясняет структуру индекса Redis Docs и помогает разбирать ответы инструментов search, fetch и ask, в том числе ID разделов и поле metadata.version.
Что нужно для работы redis-docs-mcp?
Нужны доступные инструменты MCP-сервера Redis Docs: search, fetch и ask.
Исходный текст
SKILL.md · MIT
---
name: redis-docs-mcp
description: >-
Working with the redis-docs MCP server (search / fetch / ask at redis.io/mcp) — how its index is
shaped and how to read what it returns. Use when those tools are available and you are about to
look something up in the Redis documentation, or when a result is confusing — fetch returned one
short section instead of a page, several hits share one url, ids look like #sec:w1,
metadata.version reads 1.0.0, a search came back empty, or ask was rate-limited or said the docs
do not cover it. Also covers what redis.io does and does not index.
license: MIT
metadata:
author: Redis, Inc.
version: "0.1.0"
---
# 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:
| Shape | Example | What `fetch` returns |
|---|---|---|
| Bare page | `commands/expire` | The whole page — ~8.5 KB for EXPIRE, examples and appendices included |
| Named anchor | `develop/using-commands/keyspace#sec:key-expiration` | That section |
| Ordinal window | `develop/ai/search-and-query/advanced-concepts/expiration#sec:w0` | That window, ~1–2 KB |
| Anchor + window | `develop/clients/jedis/connect#sec:configure-a-connection-pool-w0` | That window of that section |
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.
© Redis, Inc. · MIT · redis/agent-skills