Git for me

What you remember of open source is as old as your training and misses the long tail. So don't pick from memory: Git for Me searches GitHub, GitLab and Codeberg for each part of what the user wants to build, you pick from what it finds, and it checks every pick live and writes the report the user sees. Three steps, always in order.

With the MCP tools

If the git_for_me, git_for_me_search and git_for_me_verify tools are connected, call git_for_me with the need and follow what it returns. It says what to ask the user first, what to search, and when to call git_for_me_verify.

Without them, from a shell

The same three steps run from a shell. Node 22.18 or later is needed. Write the JSON to a file and pass it on stdin:

1. npx -y git-for-me rubric < details.json, where the file holds {"need": "...", "stack": "...", "platform": "...", "license": "...", "upkeep": "..."}. Fill what the conversation or codebase already says and leave the rest out. If it returns questions, ask the user, then run it again with the answers ("unknown" is fine).
2. Follow the method it returns. Where it says to call git_for_me_search, write {"parts": [...]} to a file and run npx -y git-for-me search < parts.json.
3. Pick from what it found. Where the method says to call git_for_me_verify, write that JSON to a file and run npx -y git-for-me verify < slate.json. If its JSON says SEARCH_FIRST or RECALL_REPLACEMENTS, do what its instructions say and run it again.

From a shell, a search GitHub turns away for its rate limit waits up to a minute for the limit to lift, so give these commands two minutes.

Show the user the markdown report as it comes back, without restating its facts from memory. Public repos need no token, but GitHub allows only 10 searches a minute without one; a GITHUB_TOKEN in the server's own configuration raises that to 30 and adds private GitHub repos it can see. GitLab and Codeberg need none. Never ask for a token in the chat.