Skip to content

Audit & governance

For the security or compliance owner who needs to answer: what did the agents do, who approved it, and can I prove it later?

Every governed write produces a receipt — what changed, proposed by which agent, approved by whom, when — and the audit trail is hash-chained, so tampering is evident. Read agents are structurally unable to mutate systems; irreversible actions always require human approval, enforced in code. That architecture is described in How your data is handled; this page covers the operational audit surface around it.

Activity events across the platform surfaces, including:

EventFires whenCaptured
Session initializeA coding agent opens an MCP sessionClient info, source (e.g. claude-code, cursor), team/workspace identity
Tool callAny Data Workers tool is invokedTool name, argument hash (never argument contents), latency, status
Auth failureBad, expired, or revoked credentialSource and install identity; repeated failures from one install are flagged to your onboarding engineer
Key rotationA workspace/team key is rotatedActing admin, old/new key hashes
Governed writeAn agent changes somethingThe full receipt: change, proposer, approver, timestamp

Consistent with the telemetry commitments: argument contents, query results, schemas, credentials, and PII are never in the audit stream — hashes and metadata only, except governed-write receipts, which live in your workspace.

Audit retention is 13 months by default; longer retention windows and residency requirements are agreed in your Enterprise order form rather than promised here. Exports (CSV) are available through your onboarding engineer during the pilot; self-service export in the console is part of the research-preview surface and labeled accordingly there.

Role policy is enforced server-side on every call, driven by your IdP group mapping — which is why it holds even for clients whose team config can’t express per-role rules. Client-side controls (Claude Code permission rules, Cursor team locks) layer on top for defense in depth, not as the only gate.

Ask your onboarding engineer for an audit export in week one and again before the conversion conversation — reading your own trail is the fastest way for a security team to confirm the write-safety story is real, and it’s evidence the pilot decision can cite.