Skip to content

Trust

An agent you can check, not one you have to trust

Every claim on this page is backed by a document we publish: the security reference, the DPA, the privacy policy, or the policy file that bounds what the agent may do in your repository.

Audit / task_01J2XLIVE
Issue assignedscope captured
09:41:02
Plan accepted4 files · 2 tests
09:41:11
Patch created+84 −16
09:43:48
Checks passed12 / 12
09:45:03
PR readyreceipt $1.42
09:45:07
sha256: 1d8a…9b2f · verified
Audited end to endGitHub-native workflowExisting checks stay in controlCost visible per task

If it isn't logged, it didn't happen

The audit log is written before the action, not after it, and it is append-only — the agent has no path to revise its own history. That single property is what makes the rest of this page checkable rather than reassuring: you do not have to believe that the agent stayed inside its scopes, you can read what it did, in order, with the input, the output and the reason for each step.

Least-privilege App scopes

The GitHub App requests the scopes it needs and no more. It does not ask for contents: write, administration, secrets, or any organisation-level scope.

BYOK — your provider account

Inference runs on your own Anthropic, OpenAI or Gemini key, under your own data-handling terms. We never resell tokens and we are not a reseller in the call path.

Secrets encrypted at rest

Stored credentials are encrypted at rest and decrypted in memory only at the moment of use. They are never written to logs or to audit events.

An append-only audit log

Every action is recorded with actor, repository, kind, input, output and timestamp — written before the action runs, and never updated or deleted afterwards.

The bot always discloses

Every comment states that it comes from an automated agent. There is no setting that turns this off; non-disclosure would be a bug, not a preference.

No bot-on-bot loops

CodeRabbit, Copilot, Dependabot and Renovate are recognised by identity and deferred to, so two bots never end up arguing on one thread.

The documents behind it

  • Security reference

    The exact GitHub scopes, the encryption scheme, the audit-log fields and how to report a vulnerability.

  • BYOK

    Which providers are supported, how keys are stored, and what leaves your account when the agent runs.

  • Data processing addendum

    The processor terms: what we process, on what basis, for how long, and with which sub-processors.

  • Privacy policy

    What we collect from you and from your repositories, and what we do not collect at all.

  • .maintainer.yml reference

    The file that bounds the agent: what it may touch, what always waits for a human, what it must never do.

  • Report a vulnerability

    Responsible disclosure goes to our security address, not to a public issue. We answer.

Put a fleet of AI developers to work

Bring your own keys and your own boxes, drop a .maintainer.yml, and let the loops run — audited end to end. Access is invite-only.