The trust layer for
autonomous agents.

Canary records what your agent does and blocks anything you told it not to. After each session it drafts the next rule. Five-minute install.

Get early access
See a live session
HOW CANARY WORKS

AI moves faster than we do.

Agents make millions of calls a day. Nobody's in the room when they do. Canary is. It watches the action, blocks the moves you flagged, and when something surprises you, it drafts the next rule before the next session starts.

OBSERVECONTROLBLOCKLEARN
Install

Drop in the SDK. Canary starts watching.

One npm install, one init call. Canary plugs into Claude Code, Browser Use, openClaw, Hermes — or any stack that can POST an event.

~5 minInstall to first recorded session.
See

Replay 12 hours of agent work in 30 seconds.

Every click, keystroke, and screen state — captured. Other tools trace what your agent said. Canary shows what it did.

30s12 hours of agent work, one scrubbable timeline.
Stop

Write a rule in plain English. Canary blocks it before it runs.

“Don’t touch admin settings.” “Don’t email new domains.” Canary compiles plain-English rules into policy and enforces them on every action the agent attempts.

13/13Violations caught on the ClaimDesk benchmark.
Learn

After every session, Canary writes the next rule.

Canary reviews what your agent did and drafts rules that would have prevented any mistake. One click to add — and that mistake can't happen again.

1 clickSuggested rule live before the next session.

Built for teams running

Outbound, BDR, prospecting, recycled-lead follow-up

The problem

Your outbound agent emails the wrong contact, double-sends after a retry, or fires a sequence at a closed-lost account. One flagged domain costs weeks of deliverability warm-up — and a real relationship with a real buyer.

What Canary does

Canary inspects every outbound send against your policy before it leaves your MTA. Block sends to suppression lists, enforce per-account frequency caps, require human approval above a tone-score threshold.

Example rules
  • BLOCKDon't email domains on the suppression list
  • BLOCKNo more than 2 emails per lead per 14 days
  • FLAGPause if tone_score > 0.7 — require approval
  • ALLOWAuto-approve replies to inbound threads
0 → 13wrong-lead sends caught in a 1-week production pilot
outbound-agent@photon.dev BLOCKED
Tocfo@suppressedcompany.com
SubjRe: Q4 partnership terms
Hi Sarah — following up on the proposal we sent last week. I can walk through pricing tiers today or tomorrow…
MATCHED POLICYsuppression_list · cfo@suppressedcompany.com · added 2026-03-22

Every agent running itself
needs someone watching it.

Watch. Block. Learn. Repeat.

Free tier at launch. Invite-only until then.

The magic

View source code

Click below to pull the code for this page.

Canary session log · live00:002026-04-24

Caught you.

Same action log we'd show you for a real agent session. Except the agent is you, and what you clicked was this page.

  • No events yet. Scroll up.
Events observed0
Sections visited0
Flags raised0
Attention score