Smahtiepants MCP

smahtiepants MCP: ground answers in current local documentation

Your training data has a cutoff; library, framework, language, and CLI APIs change. When the smahtiepants MCP server is connected, use its MCP tools before answering from memory for API usage, syntax, flags, defaults, examples, or current-best-practice questions.

The server searches a local DevDocs mirror with semantic + keyword search. It is read-only through MCP: use it to discover installed docsets, search excerpts, and read full Markdown pages. Do not start a shell command just to query documentation when the MCP tools are already available, and do not switch to the REST endpoint after an MCP search.

Workflow

1. Discover installed docsets

Call list_docsets first when you do not already know the exact installed slug. It returns lines like:

react: React
python~3.13: Python 3.13
css: CSS

Match the user's library or language to the canonical slug. Common language-like inputs such as python, js, ts, node, and c++ may map to canonical DevDocs slugs such as python~3.13, javascript, typescript, node, or cpp.

2. Search with scope

Call search_docs with a concise query and the most relevant slugs or languages filter.

Example arguments:

{"query": "useEffect cleanup function dependencies", "slugs": ["react"], "limit": 5}

Prefer scoped searches. Broad unscoped searches are acceptable only for initial discovery when the relevant docset is unclear.

3. Read the result and fetch full pages when needed

search_docs returns ranked text blocks with:

• <docsetSlug>:<pageId> <pageTitle>
• match kind and score
• a Read full page: hint
• a query-aware excerpt

If the excerpt is not enough to answer accurately, call get_page_content with the slug and page ID from the result.

Example arguments:

{"slug": "react", "pageId": "hooks/useeffect"}

Use startLine and endLine only when you need a specific line range from a large page. If your harness supports MCP resources, the same pages are available as smahtiepants://docsets/<slug>/pages/<pageId>.

Stay on the MCP surface for the whole workflow. Do not run smahtiepants docs page, start smahtiepants serve, or call a REST content endpoint as a substitute for get_page_content.

4. Answer from the docs

Base your final answer on what the local docs say. If they contradict your prior assumption, trust the docs and explain the current behavior. Mention the docset and page you used so the user can verify, for example:

Per the local react docs page hooks/useeffect, the cleanup function runs before re-running the effect and when the component unmounts.

If no relevant docset is installed, say that local docs are unavailable for that topic rather than guessing. If a relevant docset is installed but a specific page link is missing, get_page_content fails for the hinted page, or scoped search stays too sparse to verify current syntax, use another official documentation path available in the environment rather than retrying local variants indefinitely. Otherwise explain the limitation.

Quick reference

See references/mcp.md for the complete MCP tool surface and output details.