FailproofAI adds observability and policy enforcement to the harnesses AI agents run in. It hooks into twelve harnesses, ten coding CLIs (including Claude Code, Codex, GitHub Copilot CLI, Cursor Agent CLI, OpenCode, Pi, Factory Droid, Devin CLI, Antigravity CLI, and Goose) and two chat and assistant gateways (Hermes and OpenClaw), records every run, and blocks dangerous tool calls before they execute. It runs locally and ships 40 built-in policies.
All harnesses share one policy API and one session history. Blocking a tool call before it runs works on all twelve harnesses, and turn-end gates work on eight. Agents outside these harnesses report through a Python SDK for tracing, sessions, and audits, though enforcement there needs a hook in your own runtime. Setup wires the hooks but enables no policies until you add a policy pack, and one built-in policy that stops agents from disabling FailproofAI itself is always on.
| Policy | What it blocks |
|---|---|
| block-env-files | Reads of env and other secret files |
| warn-repeated-tool-calls | The agent looping on the same call |
| block-sudo | Privilege escalation |
| warn-destructive-sql | DROP, TRUNCATE, and unbounded DELETE |
| block-terraform / block-kubectl | Unreviewed changes to live infrastructure |
| block-rm-rf | Recursive file deletion |
| block-force-push / block-push-master | Force pushes and direct pushes to main |
Custom policies are JavaScript files committed to the repository, so the whole team picks them up, and each returns allow, deny with a message to the agent, or instruct, which lets the call through and adds context to the next prompt. A local dashboard reads the run history on your machine with no account, showing sessions, model calls, tool calls, and hook decisions, and an offline audit scans history for risky patterns and suggests policies. A hosted observability service adds fleet-wide traces, an execution graph with parallel subagents, latency percentiles, per-model cost and context tracking, SQL over traces, evaluations, scheduled audits, and alerts.
Features
- Harness hooks: one policy API across twelve coding CLIs and agent gateways
- Pre-execution blocking: stops dangerous tool calls before they run, on every supported harness
- Built-in policies: 40 policies covering secret files, loops, privilege escalation, destructive SQL, infrastructure changes, and force pushes
- Custom policies: JavaScript policies that allow, deny, or add instructions to the agent's next prompt
- Output sanitizing: reports secrets that appear in tool output
- Local dashboard: sessions, model calls, tool calls, and hook decisions from local history, with no account
- Local audit: scans run history for risky patterns and suggests policies
- Python SDK: tracing, sessions, and audits for agents outside the supported harnesses
- Hosted observability: fleet-wide traces, execution graphs, latency percentiles, cost tracking, SQL over traces, evaluations, and alerts
