Coding agent audit cockpit
A local-first audit cockpit for coding agents
ActraDeck sits in front of coding agents such as Claude Code and Codex. It is not another single-vendor dashboard: it observes sessions, redacts secrets before persistence, and keeps a replayable audit trail across agents in one local cockpit — and relays high-risk approvals where the agent supports it (Claude Code over Attach, Codex in Managed Mode; external adapters are observe-only). It is a floor against accidents around what flows through its hooks, not an OS- or container-enforced control plane.
Why sit in front of your agents?
Coding agents can run quickly and unattended. ActraDeck gives the operator one place to hold high-risk actions where the agent supports it, review what happened, and keep a vendor-neutral audit trail instead of scattering evidence across separate agent UIs — a safety net, not a sandbox against a malicious local process.
What ActraDeck does
- Approval cards for high-risk operations, with fail-safe timeout to deny.
- Secret redaction before events are persisted — native sessions are redacted on your machine by the sidecar; external adapters POST as-is and are redacted at the backend ingress floor.
- Replayable audit trails across Claude Code, Codex, and external adapters.
- Local-first operation: a sidecar and web cockpit you control.
Vendor neutral by design
ActraDeck normalizes events into one event model. Claude Code and Codex are first-class paths today, while the public ingestion contract lets other tools post normalized events into the same cockpit.