Skip to content

About

Stop being your agent's meat proxy. A local safety net for autonomous coding agents.

Resources

Security policy

Stars

2 stars

Watchers

0 watching

Forks

Repository files navigation

NoMeatProxy: stop being your agent’s meat proxy

Claude Code Hook Codex Hook OpenCode Plugin Antigravity Experimental Status: Prototype License: MIT

meat proxy · noun
A carbon-based peripheral that clicks “Allow” for an AI.

You wanted an autonomous agent. Somehow you became its full-time approval button.

NoMeatProxy is a local safety hook for coding agents running with broad permissions. It checks supported shell actions before they run and catches things like deleting uncommitted work, wiping a protected directory, or force-pushing over remote history.

“I’ll just clean that up” should not be the last thing your files hear.

Let It Cook.

Routine commands pass quietly. The ordinary check needs no second model, no network trip, and no fresh opportunity for you to click “yes, still yes”.

Look at What Could Be Lost.

File and Git checks inspect the target and local state. An ignored database still contains data. A staged edit may already have a preserved copy. The details matter.

Stop With Receipts.

A block shows the target, the risk, and the evidence. You get something concrete to review when it’s time to step in.


Quick Start • Demo • How It Works • What It Catches • Agents • FAQ


⚡ Quickstart

Requires Node.js 22+ and Git. No model subscription for the guard. No Go compiler for the standard installation. Your finger may now leave the approval button.

Install from npm:

npm install --global nomeatproxy
nomeatproxy install claude
nomeatproxy doctor claude

For prebuilt packages and other installation options, see the installation guide.

Pick your agent:

nomeatproxy install claude
nomeatproxy install codex
nomeatproxy install opencode
nomeatproxy install antigravity

Run the line for the agent you use, then restart it. Codex also requires hook review/trust through /hooks. Antigravity support remains experimental.

The installer merges into your user settings and backs up files it changes. Add --project to install only in the current project. It leaves your other hooks alone. doctor checks local setup; the agent's diagnostics confirm that it actually loaded and trusted the hook.

Take it for a spin:

nomeatproxy check "git status --short"
nomeatproxy check "rm -rf /"

These commands only inspect the quoted strings. They do not execute them. The first passes; the second is blocked with a reason.

The npm package targets macOS, Linux, and Windows. Shell inspection currently covers POSIX syntax; PowerShell and cmd.exe operations are outside its coverage. On Windows, use WSL for the POSIX workflow. Packaging tests and live agent coverage are separate; see platform details.

Changed your mind? nomeatproxy uninstall claude removes the managed hook. The setup guide covers updates, manual configuration, and the optional faster hook transport. Homebrew distribution is not published yet.

🎬 Demo

Download the MOV

Run the repeatable walkthrough from a checkout:

npm ci
npm run demo

Press Enter to advance through five real hook checks: routine work, an uncommitted edit, the same restore after staging, ignored data, and a force push. The demo uses a temporary Git repository and scripted requests; proposed commands are inspected, never executed. No live agent or published npm package is needed.

Run all scenes without pauses: npm run demo -- --auto.

The agent decides the quickest way to “fix the branch” is:

git push --force origin main

NoMeatProxy intervenes. An excerpt from the block:

NoMeatProxy stopped this tool call.

[git.remote-history] This push can rewrite or delete remote history.
  Target: "origin"
  Target: "main"
  Evidence: "Force, deletion, or mirror semantics were requested; remote ownership has not been established."

The blocked push does not run. An intentional history rewrite goes to a human for review and execution. Meanwhile, git status, searches, and a normal push can carry on.

🔎 How It Works

  1. Catch the pending shell call. An agent hook or plugin hands the command to NoMeatProxy before execution.
  2. Inspect the relevant state. Resolve deletion targets, check Git work, or preview a cleanup with a dry run.
  3. Pass quietly or block with evidence. Supported destructive actions get held when the checks find a risk or cannot establish their target.

The current release checks selected file and Git operations. It does not execute the submitted command, call another model, or send your commands to a cloud service.

🥩 What It Catches

Examples What happens
🔴 Protected paths rm targeting /, your workspace root, .git, credential directories, or paths outside the workspace Stops
🔴 Work you could lose Deleting modified, untracked, or ignored files; Git operations that would discard local work Stops with evidence
🔴 Remote history surgery Force pushes, mirror pushes, and remote branch deletion Stops for human review
🟡 Unclear deletion targets Variables, globs, or an unresolved working directory in an rm command Holds instead of guessing
🟢 Routine work Status checks, searches, normal pushes, and supported restores that preserve the affected work Passes quietly
⚪ Outside current coverage Effects inside arbitrary scripts, task runners, cloud commands, and other unsupported tools May pass without inspection

Path-specific restore and checkout checks ignore unrelated edits. Hard resets and forced branch switches are still more conservative. The design spells out the current boundaries.

🤝 Bring Your Agent

Agent Integration Current status
Claude Code Pre-tool hook Live allow/block test passed with --dangerously-skip-permissions
Codex Pre-tool hook Live allow/block test passed with approvals and sandboxing bypassed
OpenCode Plugin Live allow/block test passed with --auto
Antigravity Pre-tool hook Adapter tests pass; live validation pending

Those live tests allowed a harmless command, then blocked deletion of a disposable untracked file. They cover the tested shell path, not every action an agent can take. Setup instructions · Versions and test details

ℹ️ FAQ

Q: Will it turn every command into a committee meeting?

That’s the problem it’s here to reduce. Checks run locally. The optional fast hook path measured 9.2 ms p95 on the current benchmark, including client startup; the direct Node CLI measured 88.7 ms. Agent scheduling is extra. See the measurements and fast setup.

Q: So I can YOLO absolutely everything now?

It catches specific mistakes on supported paths. Arbitrary scripts, redirections, heredocs, Git aliases/hooks, native file-editing tools, and cloud effects are not fully covered. A pass means no implemented rule blocked the command. Some unsupported operations pass through. The roadmap tracks the gaps.

Q: Can the agent approve its own bad idea?

There is no self-approval command. For now, a human reviews and performs an intentional blocked operation outside the agent. A protected approval flow is planned.

Q: Will it ever be too cautious?

Yes. It’s a prototype. Ignored files aren’t assumed disposable, and some Git checks still inspect the whole worktree. If it blocks a legitimate cleanup, the exact command and Git state make a useful regression case. Being ignored by Git is a poor backup strategy.

Q: Can it stop a malicious agent?

No. This is an accident guard, not a sandbox. An agent with the same OS access can change its own hooks or take a path that isn’t intercepted. Keep credentials and permissions scoped to the work.

Q: Is this related to PlanSentry?

Yes. It builds on PlanSentry’s approach: inspect the consequences and show why an operation needs a human. PlanSentry integration is planned; infrastructure protection is not included in NoMeatProxy yet. Some shell parsing code is adapted from PlanSentry; its source repository is private and is not required to use NoMeatProxy. See NOTICE.

🛠️ Help Retire the Meat Proxy

Useful contributions: a command that slipped through, a harmless command that got blocked, or an agent integration we haven’t checked. Include a small reproduction with secrets removed.

npm run check       # typecheck, build, and behavior tests
npm run test:package # install and exercise the packed npm artifact
npm run evaluate    # labeled examples, including known coverage gaps
npm run bench       # measure engine and hook overhead

Design · Roadmap · Developer setup · Publishing checklist

📄 License

MIT. Shell parsing code is adapted from PlanSentry; see NOTICE.


Back to the top. No approval required.

About

Stop being your agent's meat proxy. A local safety net for autonomous coding agents.

Resources

Security policy

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages