Safety harness for agents that push to main at 3am

What it does

FailproofAI sits between your coding agent and the shell to intercept tool calls in real time. It ships with 30 built-in policies—blocking things like git push --force, recursive deletion, and API key leaks—and exposes a local dashboard to audit what was blocked and why. The whole system runs locally, so enforcement adds no network latency.

The interesting bit

The hook system is agent-agnostic, supporting Claude Code, Codex, Cursor, Copilot CLI, and roughly a dozen others without touching the agent’s core logic. Custom policies are just files dropped into .failproofai/policies/, and they can do more than block: an instruct() decision warns the agent and appends context to its next prompt, turning guardrails into guided course corrections.

Key highlights

• 30 built-in policies cover branch protection, destructive commands, and secret sanitization
• Supports at least 12 agent CLIs including Claude Code, OpenAI Codex, Cursor, and Goose
• Custom policies load automatically from .failproofai/policies/ with no flags or restarts
• Three enforcement modes: allow, deny, and instruct (warn and append context to the next prompt)
• Local dashboard at localhost:8020 logs every tool call and policy decision

Caveats

• MIT license carries a Commons Clause: free for internal and personal use, but commercial resale of the tool itself requires a separate agreement

Verdict

Worth a look if you run agentic coding tools against real repositories and want instant, offline guardrails. Skip it if you rely on cloud-hosted agents where local process hooks can’t attach.