FailproofAI/failproofai

What it solves

Failproof AI is a safety and observability layer for AI agents. It prevents agents from performing dangerous actions—such as deleting files, leaking API keys, or executing destructive SQL commands—by blocking these calls before they happen. It also provides visibility into agent behavior, allowing developers to audit runs and identify failure patterns.

How it works

The system acts as a hook or gateway between the agent and the environment. It supports twelve different "harnesses" (including coding CLIs like Claude Code and GitHub Copilot CLI, and gateways like Hermes and OpenClaw) to intercept tool calls.

Users can apply policies to these hooks. A policy can either allow() the operation, deny() it (blocking the action and sending a message back to the agent), or instruct() it (letting it through but adding context to the agent's next prompt). Policies can be built-in, custom-written in JavaScript, or installed as versioned "policy packs" from GitHub.

Who it’s for

Developers and teams deploying AI agents that have access to system tools, databases, or infrastructure, and who need to ensure these agents operate within safe boundaries without constant manual supervision.

Highlights

• Pre-execution Blocking: Stops dangerous tool calls (e.g., rm -rf, sudo, git push --force) before they execute.
• Extensive Harness Support: Works with a wide range of coding agents and gateways out of the box.
• Customizable Safety: Supports custom JS policies and shareable policy packs.
• Full Observability: Includes a local dashboard for session history and traces, as well as a hosted version for fleet-wide monitoring, latency tracking, and cost analysis.
• Audit Capabilities: Scans history for risky patterns and suggests policies to prevent them.