Skip to main content

Use These Docs with Agents

An agent reads plain text better than a web page. This site gives agents text files, a small manifest, and a plugin. The plugin adds an MCP server (a program that gives an agent tools) and a skill that tells Claude when to use it. The server reads the live site and keeps a copy on disk. So it still answers when you are offline (see Live Fetch and Cache). It has no telemetry and sends no data about you. It only requests public pages. Source: llms.txt Proposal and Model Context Protocol Specification.

Text Files​

Every build writes these files. The base URL is https://clawdipedia.cc.

FileWhat it holds
/llms.txtAn index in the llmstxt.org format. One line per page, with its title, URL, and description.
/llms-full.txtThe text of every page in one file, in sidebar order. It is large.
/<route>.mdOne page as clean Markdown. For example, /configure/status-line.md.
/llms-manifest.jsonA list of pages with route, title, description, section, md, and url, in sidebar order.
/schemas/The JSON Schemas of every config file and payload, and the Mod API type file claude-code.d.ts. See JSON Schemas.

An agent with a web fetch tool can start at llms.txt. It can pick a page and fetch its .md file.

Install the Plugin​

The plugin has the name clawdipedia-docs. Its display name is Anthropic Docs. The name clawdipedia is the marketplace.

/plugin marketplace add xsyetopz/clawdipedia
/plugin install clawdipedia-docs@clawdipedia
Breaking Rename

The plugin had the name cc-docs. Its marketplace had the name claude-code-docs. The old names are gone, with no alias. Remove the old plugin. Then install the new one. The tool names change from mcp__cc-docs__* to mcp__plugin_clawdipedia-docs_docs__*. If a permissions.allow list holds mcp__cc-docs__* (or mcp__claude-code-docs__*), change it to mcp__plugin_clawdipedia-docs_docs__*. The old rule matches nothing. So Claude asks before each call. The cache folder moved to ~/.cache/clawdipedia. The CC_DOCS_* variables are now CLAWDIPEDIA_*. So CC_DOCS_CACHE_DIR is now CLAWDIPEDIA_CACHE_DIR. The source value docs is now clawdipedia.

The plugin adds two things.

PartWhat it does
The MCP server docsGives Claude four read-only tools, such as mcp__plugin_clawdipedia-docs_docs__search_docs. It runs with plain bun. You need no bun install.
The skill clawdipedia-docsTells Claude to look up settings, environment variables, hooks, and other Claude Code details here instead of from memory. It also tells Claude which of the four sources to use. It names the source, the version, and the badge meaning in each answer.

The skill uses the MCP tools when they exist. If they do not, it fetches llms.txt and the page .md files.

Use the Command Instead​

The plugin clawdipedia-docs-cli (display name Anthropic Docs CLI) gives the same four sources as a command. It has no MCP server and adds no tool list to the context. It works only in Claude Code. It needs Bun.

/plugin install clawdipedia-docs-cli@clawdipedia

The command is clawdipedia. The Bash tool runs it. The skill clawdipedia-docs-cli tells Claude when. The package is not on the npm registry, so bunx clawdipedia-docs does not work. Outside Claude Code, clone the repository and run bun run cli/src/cli.ts.

clawdipedia search "PreToolUse timeout"
clawdipedia read configure/status-line
clawdipedia list configure --source code
clawdipedia schema settings hooks.PreToolUse

Install one of the two plugins, not both. They give the same answers.

Sources​

The first three tools take source. It is "clawdipedia" by default.

sourceWhat it readsUse it for
clawdipediaThis unofficial site.How Claude Code really works: settings, environment variables, hooks, plugins, internals, and schemas. Only this source has badges and schemas.
platformplatform.claude.com/docsThe Claude API, SDKs, models, and prompt caching.
codecode.claude.com/docsThe official Claude Code docs.
anthropicwww.anthropic.comNews, research, and legal pages. The site has no Markdown. So the server converts each HTML page.

The CLI takes the same values in --source. The old value docs is an error in both.

MCP Tools​

All four tools only read. None changes a file or your settings.

ToolWhat it does
search_docsRanks pages for a query. Gives the route, title, URL, and a snippet for each hit. It takes query, source, limit, and section. The first search loads llms-search-index.json from the site in one request. If the site has no such file, it fetches every page once. With source: "platform" or "code", it ranks pages by title, summary, path, and group from the llms.txt of that host. It does not search page text. With source: "anthropic", it ranks pages from the sitemap by the path and by a title that it derives from the path. For these three, section is a group.
read_pageGives a page as Markdown, or one H2 section of it. It takes route, source, section_heading, offset, and max_chars. A long page comes in parts. Pass offset to read the next part. With source: "platform" or "code", route is a path under /docs/en/. The tool reads only English pages. With source: "anthropic", route is a path of the site, such as research/building-effective-agents.
list_pagesLists the pages in sidebar order. Pass section to see descriptions. With another source, it lists the groups with their page counts. section lists the pages of one group.
get_schemaGives one JSON Schema of this site. With name: "types", it gives the Mod API type file. A schema is large. So you get an outline first. Pass property for one key, such as hooks.PreToolUse, a definition such as BashInput, or a type such as ToolCheckResult. If it does not have the name, it returns the closest names it has.

Every result ends with a Source: line. It names the host and says whether the text is live, from the cache, or from the cache while offline.

Use the Server with Other Clients​

The server speaks MCP over stdio. Any MCP client can run it. Clone the repo. Then point the client at the built file.

MCP client config
{
  "mcpServers": {
    "docs": {
      "command": "bun",
      "args": ["/path/to/clawdipedia/plugin/mcp/dist/server.js"]
    }
  }
}

In Claude Code, you can add it by hand with claude mcp add clawdipedia-docs -- bun /path/to/clawdipedia/plugin/mcp/dist/server.js.

Live Fetch and Cache​

The server has no bundled copy of the pages. It fetches the manifest, pages, and schemas from the site when a tool needs them. It fetches the pages of the other three sources from their hosts. The server saves each fetched file on disk. If the network fails, the server serves the saved copy. The Source: line says it is offline. If the server saved nothing yet, the tool returns an error that says the host could not be reached. The server starts without any network request.

Settings​

VariableWhat it does
CLAWDIPEDIA_BASE_URLFetches this site from a mirror instead of https://clawdipedia.cc.
PLATFORM_DOCS_BASE_URLFetches the platform docs from a mirror instead of https://platform.claude.com.
CODE_DOCS_BASE_URLFetches the official Claude Code docs from a mirror instead of https://code.claude.com.
ANTHROPIC_SITE_BASE_URLFetches the Anthropic site from a mirror instead of https://www.anthropic.com.
CLAWDIPEDIA_CACHE_DIRSets the cache folder. If it is unset, the server uses $CLAUDE_PLUGIN_DATA/cache, then $XDG_CACHE_HOME/clawdipedia, then ~/.cache/clawdipedia.
CLAWDIPEDIA_CACHE_CAP_BYTESSets the size cap of the cache in bytes. The default is 60 MB (62914560).
CLAWDIPEDIA_TTL_SCALEMultiplies the times to reuse a file in Cache Rules. The default is 1.

A base URL must use https. Plain http works only for localhost and 127.0.0.1. The server ignores any other value. It writes a warning to stderr.

Cache Rules​

RuleValue
Time to reuse a file without a requestSite manifest, platform llms.txt, and code llms.txt: 10 minutes. Site pages and schemas, and the Anthropic sitemap: 1 hour. Pages of platform, code, and anthropic: 6 hours.
After that timeThe server asks again. It sends If-None-Match or If-Modified-Since when it has a saved validator. A 304 keeps the saved copy.
Request limits8 seconds per request, 6 requests at once, at most 3 redirects, and only to the same origin. A file over its size cap (1 to 2 MB) counts as a failure.
Failed requestThe server serves the saved copy, if there is one.
404 for a page or schemaThe server deletes the saved copy. The tool says the page was not found.
404 for the manifest or llms.txtCounts as a failure. The server uses the saved copy.
Over the size capThe server deletes the least recently fetched files, pages of platform, code, and anthropic first. Then it deletes site pages and schemas. The server never deletes the index files.
Folder that cannot be writtenThe cache lives in memory until the server stops.

The server ignores any manifest entry with an odd route or a URL outside the site. It does not send or keep cookies.

Inspect Sessions and the Program​

The same marketplace has a second plugin, cc-inspect. It reads your session transcripts and the installed Claude Code binary. It makes two kinds of network request. release notes runs gh release view to get the release notes, once for each version. code read with view: "readable" runs bunx @wakaru/cli@1.14.0, once for each module. It writes only its cache of cleaned code. The cache is under CC_INSPECT_CACHE. If that is unset, it is under ./captures when that folder exists. Otherwise it is under ~/.cache/cc-inspect. code run also starts the program once, offline, with temp folders that it removes after the run.

/plugin install cc-inspect@clawdipedia

It adds the MCP server cc-inspect. This server runs with plain bun. It also adds the skill cc-inspect. The skill tells Claude when to use the server. The server has 4 tools. Each takes an action or a kind that picks what it does. Each also takes the same parameter names: version (a version or a binary path), query, name, offset, and limit.

ToolChoiceWhat it does
sessionsaction: "list"Lists the transcripts in projects/ of the config folder (CLAUDE_CONFIG_DIR, or ~/.claude), newest first. Pass project to narrow it.
sessionsaction: "read"Gives the lines of one session. Pass type to keep one line type. Pass offset to page. A session id prefix works.
sessionsaction: "search"Finds a query (text, or a regular expression with regex) in one session or in all of them.
sessionsaction: "validate"Checks each line against the transcript schema of its type, and lists the lines that do not fit. The schema is the one for Claude Code 2.1.312.
sessionsaction: "tool-calls"Counts the tool calls of assistant lines by tool name. Pass query (text, or a regular expression with regex) to keep matching names and to get up to limit sample inputs each.
codeaction: "installs"Lists the versions in ~/.local/share/claude/versions. Marks the one that ~/.local/bin/claude runs.
codeaction: "modules"Lists the files inside a binary.
codeaction: "search"Finds a query in the code of a binary, with the module and offset of each hit.
codeaction: "read"Gives a slice of one module. With view: "readable", it gives the module after Wakaru rewrote it for reading. The first call for a module runs bunx @wakaru/cli@1.14.0. It needs a network. It takes a few seconds on a large module. Later calls read the cache.
codeaction: "find"Finds code by plain words in a query, such as "sleep Monitor until-loop". It ranks the functions of a prepared version by the words in their strings, property names, and labels. It gives the top 10 with module, offset, and first lines.
codeaction: "prepare"Cleans each module of a version (stable local names, then bun fmt) and indexes its functions. It writes both into the cache.
codeaction: "run"Runs a prompt, such as /verify ARG, with claude -p in an installed version against a local fake API. It gives the user turn that the program sends, from the block that holds name (by default the command tag). It uses a fake key, a dead proxy, temp HOME and config folders, and a stub of the macOS security command on PATH. It fails when a request reaches the stub with another host or key. It also fails when the program runs longer than 60 seconds. Temp paths come back as <home>, <config-dir>, <project-dir>, <bundled-skill-dir>, and <temp-dir>. Page with offset and length.
surfacekind: "env", "setting", and the other surface kindsLists the items of one kind, each with its default, gate, module, and offset when the code shows them. With no kind, it gives each kind with its count.
surfacekind: "object-keys"Lists the keys of the object literal around an offset or an anchor text in a module. It follows spreads into the functions and modules they come from.
surfacekind: "event-keys"Lists the OpenTelemetry events of a binary. With name, it merges the attribute keys of every place the program emits that event.
surfacekind: "tool-gates"Lists the built-in tools of a binary. Each tool comes with the code in the tool list that turns it on or off. It also comes with its own isEnabled() check, when it has one. It also comes with the source of the functions they call. The list does not include the filters of each session. So a listed tool may still be off.
releaseaction: "diff"Lists the functions changed, added, and gone between two versions (from and to), each with the changed lines of its cleaned code. Pass name to keep the changes that hold a text.
releaseaction: "explain"For a name (a string, a command such as /loops, or a tool name), gives where its code is and who uses it. It also gives the gates and env reads that guard it, and whether each function is a no-op, a stub, or handled.
releaseaction: "notes"Marks each line of the release notes of version as matched (a changed function or a release:diff item holds its code span or words), gated (the match reads a tengu_ gate), or unmatched. It compares against from (else the newest older install). A match is a lead to read, not a proof.

The server also has 2 MCP prompts. They give an agent the steps of a task. release_review (with from and to) runs prepare, diff, and notes. Then it runs find and explain for each gated or unmatched line. explain_change (with name) runs find and explain for one name.

In object-keys and event-keys, a key that ends in ? is optional. A value shown as unknown comes from a computation. So its type is not known. The code tools read the code with the meriyah parser and eslint-scope. So minified names do not matter.