Stop burning agent context on whole-file dumps.

The local index your agent asks before it reads a file.

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

Works with Claude Code, Cursor, Codex CLI, Gemini CLI, Antigravity and any MCP client.

agent · dump whole files

›

▸list the repository tree1,200

▸read api/router.go14,800

▸read api/middleware.go11,400

▸read webhooks/handler.go18,900

▸read webhooks/handler_test.go9,600

▸read config/security.go6,100

✓verifySignature in webhooks/handler.go, enforced from api/middleware.go

context read62,000 tok

agent · with Archis

›

▸L1chi/search "webhook signature verification"420

↳3 symbols in 2 files

▸L1chi/get_relations verifySignature310

↳1 caller: api/middleware.go

▸L2chi/get_symbols budget 1,6001,600

↳compact projections

▸L3chi/get_files webhooks/handler.go verifySignature view=full2,400

↳the only file opened

✓verifySignature in webhooks/handler.go, enforced from api/middleware.go

context read4,730 tok

−22%tokens in a typical session

−50.6%on tasks both agents solved

0lines of your code uploaded

Transcripts illustrative. Figures measured: Claude Code, Sonnet 5 (effort high), 20 paired sessions, Archis the release after 0.49.0, early. How we measured

Measured

8 paired runs: one repository, one model, the same prompt.

An earlier Codex CLI run, priced in money. Every row, including two ties.

Measure	Reading files	With Archis	Difference	How sure
Output the model wrote	5,477	3,171	42% less	separates completely
Input tokens	190,474	102,089	46% less	ranges overlap
Cost per run	$0.289	$0.226	22% less	ranges overlap
Wall clock	224 s	157 s	30% faster	ranges overlap
Reasoning tokens	3,927	4,239	level	slightly against us
Answer accuracy	97.4%	97.5%	level	the saving is not bought with quality

Archis 0.18.0 · OpenAI Codex CLI 0.138.0 with gpt-5.5, effort high · prometheus at e6cfa720 · 8 pairs across 3 campaigns. Every figure is a mean over the pairs with its range beside it in the source; only the first row separates completely.

Index tested on 94 pinned public repositories, 538,745 files, 0 parse failures · recall ≥0.90 at a 1,600-token budget on 18 of 20 repositories.

All the tables, with sources

What leaves your machine

Your code never leaves your machine.

Parsing, indexing and agent queries run on your machine. There is no source endpoint.

No source endpoint — the API does not accept source
No repository names or file paths transmitted
Sign-in is used only to verify the licence
Indexing and MCP run on each developer’s machine, on every plan

Read privacy disclosure

Your machine (100% Local)

Source code→tree-sitter→local index→MCP stdio→Coding Agent

↳Only account licensing & team policies reach the cloud

Plans

Your repositories, or your team’s.

For a team the cost is rework: a senior engineer repairing an edit the agent made without knowing the callers.

0.76 to 0.97

Answer score, without and with Archis

5% vs 82%

Run-to-run spread in what the model writes, with Archis and without

On clone

Agent settings committed, the same for everyone

Score: Claude Code on prometheus, 20 paired sessions. Spread: Codex CLI on prometheus, Archis 0.18.0, 8 pairs. One repository each. How it was measured

Your repositories

The CLI, the local index and the MCP server
Every agent it finds, connected by chi init
All stable languages and integrations
Two active devices

Install

Your team

$10 a seat, a month

$8 if billed yearly · 3 to 200 seats

Organisation repositories
Agent settings committed in the repo, shared by the team
Seats pooled and reassigned from the portal
Policies and an audit log

See what is included

SSO and SCIM, audit streaming or an offline licence, at any team size: talk to us.

Install

Run it in the repository your agent keeps re-reading.

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