failproofai: Observe and Enforce AI Agent Runs

Overview

FailproofAI is a TypeScript and Rust tool for observability and enforcement across AI agent harnesses. It hooks into 12 coding CLIs and chat or assistant gateways, including Claude Code, Codex, Hermes, and OpenClaw, to capture runs and block risky tool calls before execution.

It is aimed at developers and teams who need both a trace of agent activity and runtime controls, rather than only reviewing outputs after a run. Local history and a dashboard work without an account; teams can also use the hosted observability service for fleet-wide traces and analysis.

Key Features

• Hooks into 12 supported harnesses with a shared policy API and session history.
• Provides built-in policies for risks such as secret-file reads, destructive SQL, privilege escalation, and force pushes.
• Supports custom JavaScript policies with allow, deny, and instruct decisions.
• Records model calls, tool calls, hook decisions, and blocked operations in run history.
• Includes a local dashboard and an offline audit command for reviewing runs and identifying risky patterns.
• Offers hosted trace analysis, execution graphs, latency and cost tracking, evaluations, audits, and alerts for teams.
• Provides a Python SDK for tracing and audits from agents that do not use a supported harness.

Use Cases

• A developer using Claude Code or another supported coding CLI can block unreviewed infrastructure changes or access to secret files.
• A team can inspect local traces to understand which model and tool calls led to an agent failure, without sending that history to a service.
• A platform team running agents across multiple harnesses can use hosted observability to compare runs and route alerts.
• A team with its own agent runtime can use the Python SDK to report sessions and audits, while implementing enforcement hooks in its own runtime.

Getting Started

npm install -g failproofai
failproofai config
failproofai policies add FailproofAI/policies
failproofai

See the README and quickstart documentation for setup details.

License in Plain Words

No standard license was detected. Without a license, the default copyright rules apply: you may read the code, but not reuse it. Check the repository or ask the authors.

Alternatives

• OrcaReplay: OrcaReplay records, replays, and forks agent runs for debugging, while FailproofAI also enforces policies that can block tool calls.
• aport-agent-guardrails: APort focuses on policy checks through framework hooks and repository guards, while FailproofAI also captures runs across coding CLIs and chat gateways.
• agentwatch: agentwatch provides local timelines, cost tracking, and alerts for coding agents, while FailproofAI can enforce policies before tool calls execute.

Considerations

FailproofAI is a young but actively maintained project: it has regular commits and releases, contributions from multiple people, and recent merged pull requests. Its contribution history is not heavily concentrated in one person. The repository data also reports no issues closed in the measured recent period, so teams should assess support and issue resolution for their own needs.

The CLI requires Node.js 20.9 or later and Rust. Enforcement capabilities vary by harness: pre-tool-call blocking is supported across the listed harnesses, while turn-end gates are available on only some. The Python SDK provides tracing, sessions, and audits for custom agents, but enforcement requires a hook in the agent's own runtime.

The README describes the license as MIT with the Commons Clause: internal and personal use are allowed, while commercial resale of FailproofAI itself requires a separate agreement. Hosted fleet observability and self-hosting options are distinct from the local dashboard, so teams should confirm which deployment and plan match their requirements.