Little Canary
Screen OpenClaw prompts with a local Little Canary service before supported agent runs
Little Canary for OpenClaw
This native OpenClaw plugin screens the current prompt at before_agent_run,
after OpenClaw builds the prompt and before supported embedded or CLI runners
submit it to a model. It sends that prompt to a separately running Little
Canary 0.4.0 service on loopback. An explicit unsafe verdict blocks the run.
Install
Use OpenClaw 2026.9.5 or 2026.9.6 on a supported Node runtime. Install and start the Python service first:
python3 -m pip install little-canary==0.4.0
ollama pull qwen2.5:1.5b
little-canary serve --mode block --canary-model qwen2.5:1.5b
Then install the OpenClaw package and grant the conversation hook access it needs:
openclaw plugins install clawhub:@hermes-labs-ai/little-canary-openclaw
openclaw plugins enable little-canary-openclaw
openclaw config set plugins.entries.little-canary-openclaw.hooks.allowConversationAccess true
openclaw plugins inspect little-canary-openclaw --runtime
Review the conversation-access grant before enabling it. The package includes
only the OpenClaw adapter; it does not install or start the Python service.
For a source checkout, use openclaw plugins install ./plugins/openclaw
instead of the ClawHub locator.
Boundary and failure behavior
Only the current event.prompt is screened. This plugin does not inspect
session history, tool results, or later text the agent reads. An explicit
safe: false service verdict blocks the run before model submission. A
service error, malformed response, or degraded coverage allows the run and
logs a warning. The plugin does not expose a fail-closed setting. OpenClaw's
isolated agent exec path bypassed this hook in the tested 2026.9.5 host and
is outside the claimed coverage.
The adapter accepts only an http://127.0.0.1:<port> service URL (default:
http://127.0.0.1:18421) and rejects redirects. It sends the prompt to that
loopback service and does not log prompt text. The service's configured canary
backend determines whether processing remains on the same machine; the
default Ollama URL is loopback. This is a sensing layer, not a security
guarantee or a replacement for tool policy and sandboxing.
Source, tests, and the host capability matrix live in the Little Canary repository.
