search-claude-docs

Search the Claude Code docs

Question to answer: $ARGUMENTS

The claude-code-docs MCP server exposes two tools. Used together they let you answer from the published docs with an exact citation while loading a few dozen lines into context instead of whole pages.

Tool	What it is	Use it for
search_claude_code_docs	Semantic search, returns page path, anchor, and a content chunk	Locate the right page for a concept or a phrasing you cannot guess
query_docs_filesystem_claude_code_docs	Stateless read-only shell (rg, grep, find, ls, tree, head, tail, sed, awk, wc, jq) over a virtual filesystem holding the docs as .mdx	Exact keywords, reading a section, dating a change

Layout of the virtual filesystem

• One directory per language. English is /en/; other languages hold translations of the same pages.
• /en/<page>.mdx mirrors https://code.claude.com/docs/en/<page>. The SDK pages live under /en/agent-sdk/.
• /en/changelog.mdx is the full release history. Each version starts with <Update label="2.1.251" description="August 28, 2026">, not a markdown heading.
• /en/whats-new/<year>-w<week>.mdx are weekly digests that group releases with context and links.

Workflow

1. Start with search_claude_code_docs when you know the concept but not the page. Read the returned paths and anchors; they are your citations. Stop here when a chunk already answers the question.
2. Switch to the filesystem tool for exact terms: an env var, a flag, a setting key, an error string, a tool name. Always give rg a path under /en/. Without a path it scans every language and floods the result with translations:

   rg -n "CLAUDE_CODE_GLOB_NO_IGNORE" /en/
   rg -n -i "bypass permissions" /en/permission-modes.mdx
   

3. Read only the lines you need. Every call is stateless, so combine in one command:

   rg -n "^## " /en/tools-reference.mdx            # table of contents
   sed -n '120,160p' /en/tools-reference.mdx        # one section
   head -60 /en/skills.mdx /en/sub-agents.mdx       # the top of several pages
   

   cat of a whole page costs hundreds of lines; use it only when the page is short (wc -l first).
4. To date a behavior, map a changelog line to its version with awk, which keeps the last Update label seen before each match:

   awk '/Update label=/{v=\$0} /Glob.*Grep|Grep.*Glob/{print v; print}' /en/changelog.mdx
   

   Then open the matching /en/whats-new/ digest for the surrounding context.
5. Quote the doc verbatim when the wording carries the answer (a version number, a default, an exact flag). Paraphrase only the surrounding explanation. End with the URL form https://code.claude.com/docs/en/<page>#<anchor> and the version numbers you found.