AI coding assistants are confident, but they answer from what they saw in training. For APIs that change quickly, that is often out of date. I kept getting Next.js caching answers that were wrong because the API had changed. When I asked where an answer came from, the assistant had no source to show me.

The usual fixes didn't suit me. Web search is fine for broad questions, but it can't see private wikis or runbooks, and it doesn't know which version you're on. Pasting docs into the prompt works for a few pages, but the Next.js docs have 845. Full RAG setups need an embedding API, a vector database, and usually an API key.

I wanted something simpler: index the docs I care about, search them on my own machine, and get back the exact sections with links. So I built Citepath.

How It Helps

Citepath gives your assistant the current docs instead of its memory. Each search returns a few ranked passages, and each one links to the page section or file lines it came from. Your assistant writes the answer from those passages and cites them, so you can check every claim.

Citepath never writes answers itself. It only finds the relevant passages.

It also cuts down how much the assistant has to read. On the Next.js docs, the top 5 passages for "how do I set the cache lifetime" come to about 4,200 characters. The five pages they come from have about 51,000 characters of text.

It works with:

• Docs websites: crawled from the sitemap or by following links, within the scope you set, and respecting robots.txt.
• Local folders: Markdown, MDX, or plain text, such as internal wikis, runbooks, or a repo's docs/ folder.

How It Works

There are three steps.

1. Add and sync a source. citepath add nextjs --url https://nextjs.org/docs/app --sync fetches the pages and splits them by heading. Each passage keeps its heading, so every citation points to the right section.
2. Search. The index is SQLite with full-text search (FTS5 with BM25 ranking), built into Node. You don't need a database server, an embedding model, or an API key. The search understands code, so getUserName, --dry-run, and "how do I configure caching" all find what you'd expect. Searches take milliseconds.
3. Connect your assistant. Citepath runs as an MCP server, so Claude Code, Cursor, and other MCP clients can search your docs directly. Your own tools can use the CLI's JSON output or a local HTTP API instead. All three return the same results.

Here's a real exchange:

Me: Using citepath, how do I set how long cached data lives in Next.js?

Claude: Add 'use cache' to the function and call cacheLife('hours'), or pass { stale, revalidate, expire }. This requires cacheComponents: true in next.config.ts. Sources: cacheLife, use cache, cacheComponents

Guardrails

• Runs locally: no account, no API keys, no telemetry. It only connects to the sources you add.
• Read-only for assistants: MCP and the API can only search and read. Adding, syncing, or deleting sources happens in the CLI.
• Indexed text is treated as untrusted: it is never executed, and agents are told not to follow instructions found in it.
• A failed sync never deletes good data: syncs are all-or-nothing. A page is removed only when there's clear evidence it's gone, such as a 404 or a deleted file. A sync that would remove a large share of a source is stopped unless you override it.
• Safe crawling: it stays inside the scope you set and refuses private and internal network addresses.

The Trade-off

Citepath matches words, not meaning, so "sign in" won't find a page that only says "authentication". In exchange, you get exact matches on identifiers, the same results every time for the same query, and nothing to pay for. Semantic search with an optional local model is something I plan to look at after 1.0.

Try It

Citepath is at version 0.1 and needs Node.js 22.16 or later. Install it, add a docs folder, and search:

npm install -g citepath

citepath add my-docs --path ~/work/project/docs --sync

citepath search "how do I configure authentication"

Then connect it to Claude Code:

claude mcp add citepath -s user -- citepath mcp

• npm: https://www.npmjs.com/package/citepath
• GitHub: https://github.com/arif-js/citepath

It's MIT licensed and open to contributions. If there's a docs site you want it to handle better, open an issue.