Agent Zero
Sign in

Open source · Apache-2.0

The autonomous engineer that never merges a guess

Agent Zero reads your pull requests, reproduces the problem, writes the fix, and proves it with a passing run. Observe mode is the default, so nothing touches your repository until you say so.

Start with observe mode Read the docs

No credit card. Runs on your own infrastructure.

agent-zero review
$ zero review --mode observe
→ mode        observe (read-only)
→ provider    anthropic/claude-opus-5
→ checkout    /srv/checkouts/acme-web@4f19c2a

✔ reproduced  src/user.ts:42  possible null dereference
✔ verified    pnpm test -- user.spec.ts  (14 passed)

2 findings · 0 writes · run 3a91f0
A review run in observe mode: findings only, no writes.

Built on the tooling your pipeline already speaks

  • GitHub
  • Git
  • Node.js
  • OpenAI
  • Anthropic
  • Containers

Everything between a review comment and a verified fix

Agent Zero is a runtime, not a chat window. Each stage is an explicit state transition you can inspect, replay, and gate.

  • Observe by default

    Every run starts read-only. Agent Zero reports what it found and what it would change; promoting to fix mode is a deliberate, auditable step.

  • One execution boundary

    Commands and file writes go through a single runner boundary with path validation and escape rejection, so there is exactly one place to audit.

  • Fixes that come with proof

    A change is only proposed once the repository's own checks pass against it. Failed verification ends the run instead of opening a hopeful pull request.

  • Any model provider

    Native OpenAI, Anthropic, and Google adapters, AI Gateway, or any OpenAI-compatible endpoint. The runtime never depends on a specific provider.

  • Leased sandboxes

    Hosted execution through a provider-neutral runner pool: every lease has a maximum lifetime, quotas are checked before provisioning, and expired sandboxes are stopped.

  • Policy you keep in git

    Modes, model selection, and permitted commands live in repository configuration, reviewed like any other change to your codebase.

What teams do with it

Agent Zero is used where a wrong automated change is expensive.

  • “We pointed it at a backlog of stale review comments and let it run in observe mode for a week. It surfaced two real null-dereference paths our linters never flagged.”
    Maya ChenStaff Engineer, Northline
  • “The part that sold us was the runner boundary. There is one file to read if you want to know everything the agent is allowed to execute.”
    Jonas WeberPlatform Lead, Latticeworks
  • “It never opens a pull request it cannot back with a green run, which means our reviewers stopped treating agent output as noise.”
    Amelia OrtizHead of Engineering, Harbor Studio

Run it yourself, or let us run it

The runtime is Apache-2.0 and always will be. Paid tiers cover hosted execution, retention, and support.

  • Community

    The full runtime, self-hosted, with your own model credentials.

    $0/ month

    • Observe, fix, and verify modes
    • Self-hosted runners
    • Bring your own model provider
    • Community support on GitHub
    Clone the repository

  • Team

    Recommended

    Hosted sandboxes and shared policy for a whole engineering org.

    $49/ month

    • Everything in Community
    • Hosted sandbox pool with leases and quotas
    • Organization-wide policy and audit log
    • 90-day run retention
    • Email support
    Start a trial

  • Enterprise

    Private deployment, procurement paperwork, and a support agreement.

    Custom

    • Everything in Team
    • Single-tenant or on-premise deployment
    • SSO and SCIM provisioning
    • Custom retention and data residency
    • Named support contact
    Talk to us

Questions people ask before the first run

Anything else, and the answer is usually in the docs.

Can Agent Zero change my repository without asking?

No. Observe mode is the default and never writes to a target repository. Fix mode has to be selected explicitly in repository policy, and every write still goes through the runner boundary.

Which model providers are supported?

OpenAI, Anthropic, and Google Generative AI natively, plus AI Gateway and any OpenAI-compatible endpoint. The provider is selected in repository policy and its credential comes from the environment.

Do I have to use the hosted tiers?

No. The runtime is Apache-2.0 licensed and fully functional self-hosted. Paid tiers exist for teams that would rather not operate the sandbox pool themselves.

What happens to my source code and secrets?

Secrets are never written to logs, prompts, fixtures, or error messages. On self-hosted runs, code never leaves your infrastructure except for the model calls you configure.

Can I cancel or change plan later?

Yes, at any time and from your own dashboard. Downgrading keeps every run you have already recorded; only retention beyond the new tier's window is dropped.

Point it at one pull request

Observe mode costs you nothing but a model call, and you will know within one run whether it earns a place in your pipeline.

Get started Talk to us