Agent runtime protection
Heartbeat monitoring, skill scanning, and a live action timeline for your OpenClaw agents. Know your agent is alive, vet every skill before it installs, and see what it actually did.
THE PROBLEM
An autonomous agent is an unattended employee.
Agents run while you sleep: they install skills, call tools, touch files, and talk to APIs. A poisoned skill or a silent crash looks identical to normal operation — until an invoice goes to the wrong place, or the agent has been down for a day and nobody noticed.
CanaryAgent adds a thin protection layer without changing your agent's core behavior. Heartbeats tell you it's alive and how busy it has been. The skill scanner catches prompt-injection payloads, backdoors, undeclared network calls, unsafe file access, and credential harvesting before installation. Every action lands in a timeline backed by CanaryShield and CanaryAudit.
HOW IT WORKS
Connect in three steps.
Copy your workspace identity
Your user ID scopes heartbeats and scans to this workspace — one click from the CanaryAgent workspace.
Add the CanaryVaults variables to your OpenClaw .env
Point the middleware at the CanaryVaults endpoints with your workspace id and API key. No changes to your agent's own code.
Drop in canary_middleware.py
Download the middleware into your OpenClaw skills folder and call start_heartbeat() when your agent boots. Heartbeats and the action timeline start immediately.
WHAT YOU GET
Everything reports into the same console.
That's the whole integration — the middleware file handles heartbeats and the action timeline from inside your skills folder. It never raises: every call returns a dict saying whether it actually reached CanaryVaults, so a plan limit or an outage cannot take your agent down. Skill scans are separate: run them from the CanaryAgent console or the skill-scan endpoint.
Does it change how my agent behaves?
No. The middleware observes and reports: heartbeats, scan verdicts, and the action timeline. Your agent's core loop, tools, and prompts stay exactly as they are — and no CanaryVaults error is ever raised into your agent, so the worst case is telemetry you did not get.
What happens when a skill fails the scan?
You get a verdict — safe, proceed with caution, or do not install — plus the specific findings: payloads, network calls, file access, credential patterns. The install decision stays yours.
FIELD NOTES
Reading that goes with this surface.
CANARYAUDIT · 7 MIN READ
Audit trails matter when AI agents can take action
If an agent writes files, calls APIs, or changes customer state, you need a record that survives dispute resolution and compliance review.
CANARYSHIELD · 4 MIN READ
Why prompt injection needs its own firewall
Prompt safety breaks when it is treated like generic moderation. Inspection has to sit before execution, not after.
Put a watchdog next to your agent.
Free tier, no credit card. Five minutes to your first alert.
Watch your first agent