Your Work, Not Your Agent's Signature: A Local Git Guardrail for AI Credits

AI coding tools are changing how we write software. They can draft a test,

explain an unfamiliar codebase, or turn a rough idea into a useful first

implementation. That part is not the problem.

The annoying part comes at the end of the workflow: a commit message or pull

request body suddenly includes an automatic line such as:

Generated with Claude Code

or:

Co-authored-by: Claude <[email protected]>

Sometimes that attribution is intentional. Sometimes it is required by a

team's policy. But sometimes it is only a default footer that slipped through

because the developer was focused on reviewing the code.

Once it reaches a commit, it is part of Git history. Once a pull request is

open, it is already on GitHub. A CI check can report the problem, but it cannot

make that publication boundary local again.

That is the problem ,

or install it with Go:

go install github.com/paladini/ai-credit-scrub/cmd/ai-credit-scrub@latest

Then, from the Git repository you want to protect, run:

ai-credit-scrub install --git

That one command installs two local hooks:

The installer does not discard an existing hook. It preserves it and chains it

before ai-credit-scrub runs. Re-run the installer if you move the binary, since

the generated hooks use its absolute path.

A practical example

Suppose an AI tool helps you implement a feature. You review the diff, test it,

and create a commit. The tool or your Git client adds this trailer:

Add account settings page



Co-authored-by: Claude <[email protected]>

With the hook installed, Git passes that temporary message file through

ai-credit-scrub before it writes the commit. The commit that exists in history

is simply:

Add account settings page

This is deliberately narrow. These ordinary lines remain untouched:

Reviewed in Cursor

Co-authored-by: Ada Lovelace <[email protected]>

Product names alone are not matches. The tool looks for complete, known agent

credit signatures for Codex, Claude Code, Cursor, Windsurf, and GitHub Copilot.

That conservative approach matters: a hygiene tool should not surprise you by

rewriting ordinary prose or human attribution.

Inspect content before changing it

The hooks are the guardrail, but the CLI is useful on its own as well. Scan a

file without modifying it:

ai-credit-scrub scan CHANGELOG.md

Clean a reviewed file in place:

ai-credit-scrub clean --in-place CHANGELOG.md

For scripts and tooling, scan and check can return JSON too:

ai-credit-scrub check --format json pull-request.md

check exits with a failure status when it finds a match, which makes it useful

when you want a local command to stop another local workflow.

Keep pull-request text on the same path

Git hooks protect commit messages, but a pull request title and body are not

Git objects. They need their own local publication path.

Use the built-in wrapper instead of calling gh pr create directly:

ai-credit-scrub pr create \

  --title "Document local hooks" \

  --body "Explain the guardrail.\n\nGenerated with Claude Code"

ai-credit-scrub cleans the title and body, then delegates to your existing,

already-authenticated GitHub CLI. It does not need an account, another token, or

a hosted backend.

You can also use a body file, which is usually nicer for a real PR:

ai-credit-scrub pr create \

  --title "Add local publication guardrails" \

  --body-file pull-request.md

Optional guidance for coding agents

Git is the universal enforcement layer because most local tools eventually call

Git. ai-credit-scrub also offers optional adapters for repositories that use

Codex, Claude Code, Cursor, Windsurf, or GitHub Copilot:

ai-credit-scrub install --adapter codex

ai-credit-scrub install --adapter claude

ai-credit-scrub install --adapter cursor

An adapter adds local guidance to use the Git hooks and the PR wrapper. It is

not magic interception, and it will not overwrite an existing agent

configuration file. If your repository already owns that file, merge the

emitted configuration deliberately.

Custom rules need a deliberate review

Every organization has its own vocabulary. You can define literal or regular

expression rules in .ai-credit-scrub.yml, but custom automatic rewriting

requires this explicit acknowledgement:

version: 1

reviewed: true

literals:

  - "Generated by Internal Coding Assistant"

The reviewed: true requirement is intentional. First run a scan, inspect what

would match, and only then allow the rule to alter text. The goal is prevention,

not an overzealous cleaner that silently changes project history.

What this tool cannot promise

Local software should be honest about its boundaries.

• 
  git push --no-verify is an intentional way to bypass client-side hooks.


• A pull request created in the GitHub web UI does not pass through local Git.


• An independent GitHub-writing MCP server can create PR text without using the
  local wrapper.

For the last case, use ai-credit-scrub pr create, or put a local sanitized MCP

proxy in front of the only GitHub-writing server available to the agent. The

v1.1.0 keeps the

implementation intentionally small: deterministic rules, local Git hooks, a

safe PR creation path, and clear limits. No surveillance layer. No model call.

No attempt to rewrite human identity.

Install it in one repository, try a commit with a known credit line, and inspect

the result. The useful outcome is simple: your published Git history says what

you intended it to say.

• Source and releases:
  


• License: MIT