Observe. Redact. Replay.
Observe sessions across agents, redact secrets before they’re stored, and keep one replayable audit trail. Relay approvals where supported: Claude Code in Attach, and Codex in Managed Mode.
No code changes. Attach to the agents you already run, locally.
For the developer running several coding agents at once — Claude Code and Codex side by side — who wants local audit, redaction, and guardrails on their own machine.
A real, read-only walkthrough of the running cockpit — cross-vendor sessions, a live approval denied on camera, redaction, and audit. Nothing is mocked.
A destructive command shouldn’t run before a human sees it. A rm -rf is flagged high-risk and blocked until you allow or deny — and auto-denies if no one answers in time. (Claude Code relays over Attach, the default; Codex relay uses Managed Mode.)
Detected keys, tokens and .env values are masked before anything is stored or shown. On the agent path the sidecar redacts them before it persists or forwards an event; and every event — including ones an external adapter POSTs straight to the ingestion API — passes an unconditional redaction floor at the backend before it is stored. Detected secret patterns are masked before persistence; the cockpit shows redaction markers and a per-kind count (high-entropy secret, credential assignment, URL credential …). Detection is best-effort pattern matching (gitleaks-style rules + custom regexes) — a strong safety net, not an absolute guarantee.
When the codebase is your client’s, you want a clear record of what your agents did on it. Every observed session lands in one audit log — approval decisions, redaction counts, high-risk ops. Filter by repo, export JSON or CSV to hand over, and replay any session step by step.
For contractors & client work
Claude Code and Codex on one wall — each lane shows state, current action, and the repo@branch it’s in. Stalls are flagged with evidence (process, event, stdout, file, model-stream), so you can tell at a glance which session actually needs you.
For multi-session power users
Attach to the agents you already run — no code changes, your data stays on your machine.
./scripts/quickstartOne command: local backend, web UI, and Claude Code & Codex attached.
cd ~/project && claude # or: codexRun them as usual — no code changes, no new flags.
open http://localhost:55400Watch every session live across both agents, and replay the full audit trail.
Two ways to connect. Attach is the zero-friction default; Managed Mode owns the process for full control.
| Capability |
Claude Code Attach (default) |
Codex Attach (default) |
Managed Mode |
|---|---|---|---|
| Live observation & audit | ✓ | ✓ | ✓ |
| Secret redaction before persist | ✓ | ✓ | ✓ |
| Approval relay (allow / deny) | ✓ | Managed only | ✓ |
| Replay | ✓ | ✓ | ✓ |
Codex over Attach is fully observed and audited in the same audit trail; its approval requests appear — and can be allowed or denied from the cockpit — only in Managed Mode.
Open-source, local-first, vendor-neutral. Point it at the agents you already run — no code changes, and your code and events stay on your machine.