Failproof AI
A local policy layer that intercepts risky coding-agent actions before they become incidents.
What does this agent do, and when should you use it?
Failproof AI is a local runtime policy tool for coding-agent CLIs that operates through hooks around agent tool calls. The repository explicitly lists support for Claude Code, OpenAI Codex, GitHub Copilot CLI, Cursor Agent CLI, OpenCode, Pi, and other named CLIs. Its installation flow activates 30 built-in policies and exposes a local dashboard at localhost:8020. Documented protections include direct pushes to main/master, force pushes, commits or history operations on the primary branch, recursive deletion, and API-key exposure in agent context. Teams can add JavaScript policies under `.failproofai/policies/` and commit them for teammates to load after pulling.
After failproofai runs, it hooks into supported agent CLIs and evaluates their tool calls. Built-in policies inspect actions such as git push --force, direct pushes to main/master, commits/merges/rebases on main/master, recursive file deletion, and API keys entering agent context; policies return allow(), deny(message), or instruct(message). Custom JavaScript policies are registered with customPolicies.add(), can match events such as PreToolUse, and can inspect tool input through ctx.toolInput, including file_path. Each agent tool call is logged locally, and the dashboard shows what ran, what was blocked, and the policy message returned to the agent.
- A developer using Claude Code or Codex on a production repository who wants direct pushes to main/master stopped before execution.
- An engineering team that permits agent-driven Git workflows but needs to block `git push --force` and primary-branch commits, merges, or rebases.
- A maintainer assigning cleanup work to a coding agent who wants recursive deletion intercepted by policy.
- A security-conscious team that wants `sanitize-api-keys` to prevent API keys from leaking into agent context.
- A repository owner who wants to block writes to production paths with a committed `no-production-writes` policy.
What are this agent's strengths and limitations?
- Applies policy at the agent tool-call boundary, allowing it to block risky Git actions and recursive deletion before they execute.
- Ships with 30 built-in policies, with explicitly documented coverage for primary-branch safety, force pushes, and API-key context exposure.
- Supports repository-level JavaScript policies that autoload from `.failproofai/policies/`.
- Keeps a local tool-call record and provides a localhost dashboard for executions, blocks, and policy feedback.
- The documented integrations are named agent CLIs; the supplied material does not describe a generic protocol or integration path for arbitrary agent runtimes.
- Adoption requires shell usage, npm, and local filesystem access rather than a hosted service workflow.
- Commercial resale of Failproof AI itself requires a separate agreement under its MIT plus Commons Clause licensing.
- The supplied material links to, but does not detail, the full built-in-policy behavior, false-positive handling, or performance measurements.
How do you install or deploy this agent?
No credentials are documented. Install globally with npm and install the policies:npm install -g failproofaifailproofai policies --install
Then start it:failproofai
The first run presents a prompt; set FAILPROOFAI_NO_FIRST_RUN=1 to disable it.
How do you use this agent?
Run failproofai after installation; the README states that 30 built-in policies activate immediately. Open localhost:8020 to inspect local session and policy activity. For repository-specific controls, add a JavaScript file under .failproofai/policies/ and register it with customPolicies.add({ name, match: { events: ["PreToolUse"] }, fn }); have fn return allow(), deny(message), or instruct(message). Commit the policy so teammates can load it on their next pull.
How does this agent compare with similar options?
Failproof AI is not presented as a replacement for Claude Code or Codex; it adds a hook-based runtime policy layer to those coding agents and other listed CLIs, including GitHub Copilot CLI, Cursor Agent CLI, OpenCode, Pi, Hermes, OpenClaw, Factory Droid, Devin CLI, Antigravity CLI, and Goose.