When your AI agent needs a chaperone, a bouncer, and a snitch

What it does ClawKeeper hardens OpenClaw autonomous agents through three concurrent mechanisms. Skills inject security policies directly into the agent context; a plugin audits configurations and monitors behavior at runtime; and a decoupled watcher process oversees state evolution, ready to halt high-risk actions or demand human confirmation without coupling to the agent’s internal logic.

The interesting bit The project treats safety as a separation-of-powers exercise. By keeping the watcher independent and system-agnostic, it attempts to guarantee that a rogue or misdirected agent cannot disable its own chaperone. The README bills it as “The Norton for OpenClaw,” which is either a warning or a promise depending on your nostalgia for 1990s antivirus pop-ups.

Key highlights

• Three complementary layers: instruction-level skills, runtime plugin enforcement, and independent watcher middleware.
• v1.1 ships concrete guards: exec-gate blocks destructive shell commands via regex, path-guard locks sensitive files, budget-guard enforces token limits, and a permission store with HMAC-SHA256 integrity protection tracks allow/deny decisions.
• The watcher runs locally or in the cloud and is described as deployable across personal, enterprise, and intranet environments.
• A self-constructed benchmark of seven adversarial categories is cited, though the README offers no specific metrics—only a chart and the claim of “optimal defense performance.”

Caveats

• Benchmark claims are unquantified in the README; no scores, margins, or statistical detail accompany the “optimal” assertion.
• Despite the watcher’s advertised system-agnostic design, every integration example and command reference targets OpenClaw exclusively.

Verdict Worth evaluating if you are running OpenClaw agents in production and want defense-in-depth without betting everything on a single gate. If you are not in the OpenClaw ecosystem, treat this as architecture inspiration rather than a drop-in solution.