GuardClawAI Agent Safety Monitor

Risk-scores every tool call before execution. Blocks dangerous ones. Stays out of the way on the rest.

🛡️

Real-time Risk Scoring

Every tool call scored 1–10 by a local or cloud LLM judge before execution. Safe ops get a fast pass. Dangerous ones get blocked.

⚡

Zero Friction on Safe Ops

Built-in fast paths skip the LLM for git, npm, read-only commands, and other obviously safe operations. No added latency for 80% of tool calls.

🔒

100% Local by Default

Runs entirely on your machine using LM Studio, Ollama, or the built-in MLX engine. No cloud account needed. Your code never leaves.

🤖

Works with 7 Agents

Claude Code, Codex CLI, Gemini CLI, OpenCode, OpenClaw, Cursor, and GitHub Copilot CLI — all supported out of the box.

🧠

Learns from You

Adaptive memory records approve/deny decisions as evidence. It helps calibrate the judge without blindly trusting compressed patterns.

🔗

Multi-step Attack Detection

Chain analysis tracks sequences across tool calls — catching attacks like read ~/.ssh/id_rsa → curl evil.com that look harmless in isolation.

We'd Love Your Feedback

GuardClaw is early and we're actively listening. Whether you hit a bug, want a feature, got stuck on setup, or just want to say something works well — open a GitHub issue. Every piece of feedback shapes the next release.

No template, no sign-up hoops — just tell us what's on your mind at github.com/TobyGE/GuardClaw/issues.