I Built a Tool That Gives Your GitHub a Savage Roast — and a Real Score

If you've spent any time browsing GitHub profiles, you know the problem: stars are bought, followers are farmed, and PR counts can be padded with one-line typo fixes. A profile with 5,000 followers and 3,000 stars might look impressive — until you realize half those stars came from mutual-starring rings and most of those contributions were self-merged README edits.

So I built ghfind. It gives every GitHub account a 0-to-100 score across six weighted dimensions, then — if you ask for it — writes a brutally honest roast of what it finds.

Try it on yourself first. That's what everyone does.

The Score That Actually Means Something

Here's the thing most GitHub metrics get wrong: they measure popularity, not value. Stars tell you if a project is well-known. Followers tell you if someone has reach. Neither tells you whether the person behind the account does meaningful work.

ghfind scores six dimensions:

• Account maturity — longevity, not just recency

• Original project quality — repos you built, not forks you hoarded

• Contribution quality — the depth of your PRs and commits, not the count

• Ecosystem impact — whether your work gets used and depended on

• Community influence — genuine engagement, not follow-back loops

• Activity authenticity — minus penalties for patterns that look farmed

The engine is fully deterministic. Same inputs, same score, every time. No LLM touches the number. The scoring core is open-sourced as the github-account-value skill under AGPL, with the logic ported line-for-line from Python to TypeScript and locked by unit tests. You can audit it, reproduce it, or run it yourself.

The roast, though? That part uses AI, and it does not hold back.

Exposing the Farmers

This is the part that surprised me while building it. Out of nearly 19,000 scored accounts, fewer than 1% triggered the strongest red flags — but when they did, the patterns were unmistakable.

Near-identical PR titles across dozens of repos. Contributions clustered entirely on repositories the account doesn't own. Suspicious star and follower inflation curves that don't match any organic growth pattern. Template repos pushed with no meaningful changes. Fork-hoarding disguised as portfolio depth.

The engine catches what raw numbers hide. A high PR count with zero original repos and templated titles isn't a productive developer — it's a pattern. ghfind calls it out.

What makes this fair is that the score only uses public data. Private organization contributions are invisible, which the site is upfront about. A low score isn't a judgment on a person — it's a description of a public footprint. And only accounts scoring 60 or above appear on the public leaderboard, so there's no harassment vector.

The Discovery Angle Nobody Talks About

Scoring is fun. The roast is funnier. But the part of ghfind I think has the most long-term value is the discovery layer.

The leaderboard is a ranked directory of genuinely strong open-source contributors, filterable by programming language and organization. You can search for developers in your ecosystem, compare two accounts head-to-head, and find people worth collaborating with — without wading through inflated profiles.

There's also a percentile system: every scan tells you what share of scored developers you beat. It turns "am I good enough?" into a concrete, checkable number with a breakdown showing exactly which dimensions are holding the score down.

For the developer being discovered: you can embed a live score badge or a developer card in your README or portfolio. It updates as your public work grows.

Built for Developers, Not Just Visitors

The entire site is API-first. Everything the web UI does is available programmatically:

• REST API — full OpenAPI spec at /openapi.json

• npm and PyPI packages plus a CLI (ghfind scan hikariming)

• MCP server at /mcp — plug it into Claude, Cursor, or any MCP-compatible agent with score_user, scan_user, compare_users, get_leaderboard, and search_users tools

You can bring your own LLM key for the roast. It lives in your browser's localStorage, never touches the server, and works with any OpenAI-compatible API. Want to use Groq or a local model? Point it at your endpoint and go.

The whole thing is self-hostable. Clone the repo, set a GitHub token and an LLM key, deploy to Vercel in a few minutes. AGPL for the code, trademark reserved for the name and branding.

Try It

Go to ghfind.com, type in your GitHub username, and see what happens. The scan takes about 30 seconds. You'll get a score, a tier from Trash to God, a dimension-by-dimension breakdown, and the option to generate a roast.

Then type in some usernames you admire — or some you're suspicious of. That's where it gets interesting.

The repo is at github.com/hikariming/ghfind. Stars, issues, and contributions are welcome. If you find a scoring edge case or want a language leaderboard that doesn't exist yet, open an issue.