Replay what your agents did, step by step.

• Watches, never controls

  There is no pause, no kill and no approve anywhere in the console. It reads what your agents report after the fact, which means it can tell you a rule was broken and could not have prevented it.

• Runtime attribution

  Claude Code reports per-step tokens and tool arguments. Codex reports commands, exit codes and per-turn tokens. Anything else reports run totals only, and the console says which rather than showing blank columns.

• Advisory policies

  Rules are checked after a run finishes and never stop one. A budget sends an alert and the agents keep working, because losing the record of the runs that caused an overspend is the wrong trade.

• One queue

  Findings raised by you, by an AI reviewer, or by a policy check all land in the same brief. Yours are marked raised by a human, because that carries different weight.

await fetch('https://review-my-site.com/api/agents/ingest', {
  method: 'POST',
  headers: { authorization: 'Bearer agk_...', 'content-type': 'application/json' },
  body: JSON.stringify({
    run: { id: runId, agent: 'support-bot', environment: 'prod' },
    events: [
      { seq: 1, kind: 'start' },
      { seq: 2, kind: 'model_call', name: 'plan', model: 'claude-sonnet-5',
        input_tokens: usage.input_tokens, output_tokens: usage.output_tokens },
      { seq: 3, kind: 'end', detail: { status: 'succeeded' } },
    ],
  }),
});

Watch takes 100,000 events a month. Audit raises that to 500,000, and Govern to 2,000,000.

Free is the absence of a subscription, not a trial. It does not expire, nothing auto-converts, and doing nothing charges nothing. Every figure above is read from the same catalogue the checkout charges against, so a price shown here cannot differ from a price taken.

• The lanes

  One per agent, with every step positioned by when it started and how long it took. Steps overlap where agents worked at the same time, and a lane sitting empty means it was waiting.

• The diff panel

  Walks backwards to the most recent step that wrote something, so it holds the current file rather than emptying between writes. Before the first write it says so.

• Spent and tokens

  Cumulative sums across every step that had started, so at the end of the run they equal the totals on the outcome panel exactly.

• It refuses

  Ask the copilot something outside the run data and it says it does not have that, rather than something plausible. The refusal is the default answer, not the last one it tries.

• It cannot act

  The copilot cannot start a run, stop one, change a policy or a budget, or read source outside a recorded diff. That list is on the screen, and it is read from the code that does the grounding.

• Seven reviewers, five on

  Each publishes its own hit rate, because a reviewer you disagree with is one you need to be able to switch off, and you cannot judge that without the number.

• One drawer

  Every entry point opens the same one with the right step in it: the replay, the live run, a token anomaly, or the action row on an AI summary.

• No duplicates

  Flagging the same step twice confirms the finding that already exists rather than raising a second one, because two descriptions of the same nine seconds are still one problem.