Agents declare a plan. 05 holds them to it, checks the result on the exact code and signs the evidence.
You approve by exception, at your terminal or from your phone.
Website · Get started · Guide · Your own console · CI gate
An agent can write fifty thousand lines in an afternoon. Nobody can review that, so pull requests get approved unread, or people go back to writing the code by hand. 05 replaces the review with proof. You own the request and the acceptance. Agents own the plan and the work. 05 holds them to the plan. When everything matched, there is nothing to read. When something did not, you read one page, not the diff.
- Agents held to their plan. Each agent declares what it will do and where before its first edit. 05 seals that plan where the agent cannot change it, and the agent's hooks refuse what your rules forbid.
- Evidence, not "tests pass". Your checks run on the exact files, and the result is signed as a receipt anyone can verify offline. No model judges anything; the gate is fixed rules.
- One page instead of the diff. The brief compares the plan with what changed and what was proved. A 500-file build still fits on one page.
- Several agents, one codebase. Each claims an area; a clash is caught at the plan and at every edit, and changes land one at a time.
- Approve from your phone. Run your own console: see every agent's work, approve with a passkey, sign off that you tested a change, throw away abandoned work. Self-hosted, one container, encrypted end to end.
- Your rules, honestly enforced. Protected, locked and append-only paths, testing before a push, an organisation floor. Hooks are advisory, a required CI check enforces, and 05 says which is which.
It works with the agents you already run (Claude Code, Codex, and any agent with hooks). It is not an agent, not a code reviewer powered by a model, not a Git host. Free and open source; everything runs on your machine, with no account and no telemetry.
| # | Principle | What it means |
|---|---|---|
| 01 | Policy | Your rules in .05/policy.toml, set by answering a few questions. |
| 02 | Gates | Where the rules hold: the agent's hooks, a pre-push gate, a required CI check. |
| 03 | Intent | The agent declares its plan before it starts; 05 seals it and holds the agent to it. |
| 04 | Evidence | Checks run on the change itself, and the result is signed. |
| 05 | Review | Only by exception: one page, never the diff. |
curl -fsSL https://use05.dev/install.sh | sh # or: npm install -g @use05/cli
cd your-repo
use05 init # a few questions, then rules and hooks
git add .05 AGENTS.md && git commit -m "Set up 05" # your rules count once committed (add .codex too if init wrote it)
use05 key new --approver # your approval key, behind a passphrase no agent hasThen work with Claude Code or Codex as you do now. When an agent is done:
use05 brief # one page: what was asked, what changed, what was proved
use05 approve --outcome # only when the brief asks for a person
use05 land # the agent's branch lands on main and is cleaned up
git push origin main # checked again on the way out; no pull requestmacOS and Linux, x86_64 and arm64. The guide covers the rest, including working without an agent.
 |
 |
| A 497-file build, one page. What changed by area, what fell outside the plan, what 05 refused and why. | Prove, then verify. A check on the exact tree, signed, verifiable offline. |
 |
 |
| Land it, and the agent's branch and worktree are cleaned up. | Agents are held to the rules by their own hooks. |
- Using 05: how it works, set up, working with agents, the brief, approvals, every setting and command
- Your own console: approvals from your phone, self-hosted and encrypted end to end
- Turn on the CI gate: the control that enforces
- Receipt format · Extending 05 · AI governance mapping
- Security policy · Contributing · Changelog
Apache-2.0. See LICENSE.