01 / CHANGE CONTROL

Change control for code written by AI agents.

Mandate's recorder runs inside Claude Code in your repository. It records each session: the instruction, the tool calls, the commits and the checks the agent ran. Credentials are stripped on your machine and again when they arrive. GitHub tells Mandate what landed and what was deployed, and each change is judged against the mandate in force at the time.

STATEWHAT IT MEANSHOW IT'S WORKED OUTWHAT HAPPENS
WITHIN MANDATEA mandate a person granted covered it when it landedThe mandate's scope against the files and commands in the changeWaits for sign-off
OUTSIDE MANDATEIt touched something the mandate doesn't allow, such as migrations, dependencies or deploymentRead from the change and from the commands the session ranEach breach named on the change
NO MANDATENothing authorised it when it landedA mandate is dated when it's granted and can't be backdatedNamed as an exception
CHECKEDTests ran in the session, or in CISession checks are the agent's own account; CI results are observedLabelled self-reported or observed

02 / THE SESSION

Who asked for what, and what the agent did.

Each change links to the sessions that made it. A person reviews the change and its exceptions, then signs it off, and the sign-off is sealed into the month's change control report.

  • The instruction, the closing summary and a hash of the transcript
  • Whether a person was directing the session or the agent had full delegation
  • A review counts as independent only if a separate session did it after the change
  • The production deployment matched to the commit that shipped it
One change in Mandate: its labels, the mandate that covered it, the agent session with its instruction, the checks and the reviews.
One change, and the session behind it

03 / ASSISTANTS

Let an AI assistant draft the work. It can't sign it.

Claude, Cowork, ChatGPT, Codex and Cursor can connect to Mandate, with your consent, through its MCP server. An assistant works for one person in one organisation. It reads your registers and documents and drafts entries through the same checks the forms use.

Everything it writes is marked as drafted by an AI agent. It's refused every act that seals: it can't approve a document, accept a risk, approve a supplier or mark a control implemented.

ConnectOAuth sign-in, or a key you can revoke
ReadRegisters, documents, the Statement of Applicability
DraftNew entries and edits, marked as the agent's
SealNever. That's a person's act

04 / AI IN THE PRODUCT

Where Mandate uses AI itself.

  • Suggested entriesDrawn from your set-up answers and your other registers. You edit them and tick the ones to keep.
  • Reading imported documentsA first guess at a document's title, type, owner and the clauses and controls it covers. You confirm it, and Mandate shows where the document itself says otherwise.
  • Drafting an answerIn a guide, a draft for one field from what you've already entered. You edit it before it's saved.
  • Switched off per organisationAn owner can turn it off, and every guide still works without it.
  • What's sentCredentials are stripped first, and only a hash and a size are logged.

05 / LIMITS

What the recorder doesn't do.

  • It doesn't block the agentIt records and never stops a session. Authority is checked when the change is judged, not by getting in the way.
  • It only sees sessions it's installed inA change made without the recorder still arrives from GitHub, marked as unattributed.
  • It doesn't decideMandate labels each change. A person signs it off.

AGENTS IN YOUR REPOSITORIES?

See a recorded session, from instruction to sign-off.

We'll show you a change made by Claude Code under a mandate, and one made outside it.

Request a walkthrough