# Archis

> A local index your coding agent asks before it reads a file. In its early benchmark with Claude Code, agents used about 22% fewer tokens in a typical session. Your code never leaves your machine.

- Website: https://archis.dev
- Case study: https://yausell.com/work/archis
- Stack: Rust, Go, React, Vercel, MongoDB, Coolify

## What it is

A local index of a code repository that AI coding agents query before they read whole files. It runs as an MCP server with search, symbol, relation and file-view tools, and works with Claude Code, Cursor, Codex CLI, Gemini CLI and any MCP client.

```
curl -fsSL https://archis.dev/install.sh | sh && chi init
```

## Privacy model

Parsing, indexing and agent queries run on the developer's machine. No source code, repository name or file path is sent to the service; sign-in only verifies the licence.

## Results so far

The published benchmark is labelled early. With Claude Code over 20 paired sessions, agents used about 22% fewer tokens in a typical session and about 50.6% fewer on tasks both runs solved. The method and full tables are public, linked under sources.

## How it's built

A CLI and index written in Rust and Go, with a React web front end on Vercel. Personal repositories are free; teams pay per seat.

## Receipts

- Fewer tokens, typical session: about 22% (Archis early benchmark, Claude Code, 20 paired sessions, 2026-11-01, https://archis.dev/verify#benchmark)

## Sources

- [archis.dev](https://archis.dev)
- [Benchmark and method](https://archis.dev/verify#benchmark)
- [Privacy](https://archis.dev/privacy)