Open source · MIT · v0.1

Let AI write code. Make your checks approve shipping.

CodeGate is a local quality gate for Claude Code, Codex and plain Git. Before an agent commits, pushes or opens a PR, it runs your tests, lint and type checks. If anything fails, the command is blocked and the agent is told what to fix.

Install the plugin View on GitHub

● Claude wants to run git commit -m "add totals"
  PreToolUse · CodeGate

OK   merge conflicts
OK   python syntax
OK   secret patterns
FAIL unit tests
      AssertionError: -1 != 5

BLOCKED Fix the failures, then rerun
        the same action.

● Claude fixes add() and retries
PASS 4 checks · commit allowed

How it works

There is no AI reviewer and no API key. CodeGate only runs the checks your repository declares in codegate.toml, and it enforces them in three places.

Agent proposes

Claude or Codex wants to run git commit, git push, gh pr create or a deploy command you listed.

Hook intercepts

The plugin's PreToolUse hook runs conflict, syntax and secret scans plus your configured commands.

Pass or block

On pass, the normal permission flow continues. On fail, the command is denied with the failing output.

Git and CI recheck

pre-commit and pre-push hooks gate ordinary Git, and the same runner can be a required CI check.

Install

The plugins load in every project, but they only act in repositories you opted in on your machine: a codegate.toml plus a one-time init. Everywhere else, commands pass through untouched. Requires Python 3.11+ on macOS or Linux.

# inside Claude Code
/plugin marketplace add Relay-harness/codegate
/plugin install codegate@codegate

# then, in the repo you want to protect
/codegate:setup

Add the marketplace and install. This adds a Bash hook, an MCP tool (codegate_check) and two skills.
Run /codegate:setup in your repo. It detects your test and lint scripts, writes codegate.toml and installs Git hooks.
Review and commit codegate.toml and .codegate/runner.py. Use /codegate:check any time.

# add the marketplace
codex plugin marketplace add Relay-harness/codegate

# restart the ChatGPT desktop app, open Plugins,
# install "codegate", then trust its hook in /hooks

# in the repo you want to protect
uvx --from git+https://github.com/Relay-harness/codegate.git \
  codegate init --git

Add the marketplace, then install CodeGate from the Plugins directory in the ChatGPT desktop app (Codex).
Trust the hook with /hooks. Codex does not run plugin hooks until you review them.
Opt the repo in by asking Codex to use the CodeGate setup skill, or by running init yourself. Its codegate_check MCP tool takes the repository path.

Works in local Codex threads, where tools can reach your files. The ChatGPT web app cannot run local tests, so CodeGate cannot gate code written there. Use the Git hooks and required CI for that.

uv tool install git+https://github.com/Relay-harness/codegate.git

# in your repository
codegate init --claude --git
codegate doctor
codegate check

# in CI
python3 .codegate/runner.py check --json

No plugin needed. Install the CLI and add Git hooks to any repository.
Teammates bootstrap with python3 .codegate/runner.py install --git after cloning.
Make it binding by running the runner as a required status check with branch protection.

One file of policy

Commands are argv arrays with no implicit shell. They run in order, stop at the first failure, and each has a time limit. With no checks configured, shipping stays blocked. An empty policy never counts as a pass.

# codegate.toml
version = 1
timeout = 300
secret_scan = true
ship_commands = ["npm publish", "vercel --prod"]

[[checks]]
name = "tests"
command = ["npm", "run", "test"]
timeout = 120

Honest about limits

A pass means your declared checks passed on the code that was checked. It does not mean the code is bug-free.

What it does

Blocks commit, push and PR commands from agents when checks fail
Gates ordinary git commit and git push through hooks
Detects tests that modify source files during a run
Never caches a pass as permission to ship later
Runs locally with zero runtime dependencies

What it doesn't

Act as a sandbox: checks run with your user permissions
Parse every possible shell trick: aliases and wrapper scripts can get past the hook
Stop someone who deliberately bypasses local hooks: use required CI
Control the ChatGPT web app or code copied from it
Replace code review or a real security scanner