Skip to main content
agentsearch mounts any documentation site as a filesystem your agent can navigate with the tools it already knows: tree, grep, cat, find. No API key, no account, no install — one npx command and the docs are a directory you (or your agent) can cd into.
Why a filesystem? Agents were pre-trained on Unix. They already know cat README.md and grep -r "auth" .. Giving them a filesystem instead of yet another MCP tool means zero schema overhead, zero context tax, and no new abstractions to learn.

​
Quick start

​
Browse any docs site

That’s the entire interface. The first call indexes the site (cold index ~30–120s, then ~2s on subsequent visits and ~100ms when locally cached). You’re dropped into a bash shell scoped to that site:
All standard Unix tools work in-process: tree, ls, cat, grep, find, head, tail, wc, cd, pipes, aliases. The shell starts in the docs root — use . for relative paths.

​
One-shot mode (for agents and scripts)

Use -c to run a single command and exit. This is the mode most agents use:

​
Wire it into your agent

agentsearch ships with two helpers — setup and agents — that generate an agent-ready instruction snippet for any URL. Pipe the snippet into your agent or append it to its rules file.

​
Claude Code

​
Codex / OpenAI

​
Append to AGENTS.md / CLAUDE.md / .cursorrules

​
OpenCode

​
Gemini CLI

​
GitHub Copilot CLI

Works with anything that reads an instructions file or supports skills — Claude Code, Cursor, Copilot, Codex, Gemini CLI, OpenCode, and more. After setup, the agent receives a short snippet like this:
When the agent needs to verify how something works, it runs the command, gets the actual current documentation, and writes code against that — not against whatever was in its training data.

​
How it works

Three pieces, no magic.

​
1. Index

The backend crawls the site once. It respects llms.txt, auto-detects OpenAPI specs (which appear under /api-spec/ in the filesystem), and normalizes URL paths so the filesystem mirrors how you think about the docs, not how the URLs are structured. For example, https://better-auth.com/docs/installation becomes /installation.md, not /docs/installation.md. Indexes are namespaced by canonical site ID and shared across all users — index docs.stripe.com once, everyone benefits. Public docs only, unauthenticated by design.

​
2. Serve

The backend exposes filesystem operations as HTTP endpoints (load, read, grep, ls, tree, find). Everything is gzip-compressed. The CLI maintains a disk cache at ~/.cache/nia-docs/ keyed by namespace and indexed_at timestamp. Backend cache TTL is 5 days.

​
3. Shell

The shell runs on the client, not in a container or VM. agentsearch uses just-bash — a TypeScript bash reimplementation that supports grep, cat, ls, find, cd, tree, pipes, and aliases. The whole filesystem is an in-memory JavaScript object, so a grep -r "webhook" . over 500 pages completes in milliseconds because it’s pure string matching in memory. Result:
  • ~100ms to boot when locally cached
  • ~2s when the site is already indexed on the backend
  • ~30–120s for a cold index of a brand-new site
  • Zero per-session compute on the server

​
Telemetry

Every command run inside a shell session is logged: command name, success/failure, duration, file count for greps. Not the content — just the patterns. We use this to understand how agents navigate docs so we can keep making the filesystem better. Opt out:

​
When to use agentsearch vs. other Nia products

agentsearch is the lightest possible front door to Nia — public docs, zero state. Graduate to the full platform when you need persistent indexes, private sources, or research workflows.

agentsearch.sh

Manifesto, live demo, and source

GitHub

nia-docs source code
Need help? Join our Discord community for support.