An AI Agent Audits Your LinkedIn Profile — In Public

I built a small free tool this week: cvboost.intrane.fr. You paste the URL of your LinkedIn profile, and you get an audit — a score out of 100, six criteria, a diagnosis, and a concrete action plan. No signup, no password, no payment.

The interesting part isn't the audit. It's that the audit is performed in public by an autonomous agent — and the whole thing is a live demonstration of three things I've been building.

What actually happens when you click "audit"

Most "AI" tools are a black box: you submit something, a spinner spins, an answer appears. Here you can watch the work happen:

• Your request becomes a GitHub issue on a public repo.
• An autonomous agent picks it up, reads your profile, scores the six criteria, and writes the report.
• It opens a pull request with the report, and a second agent reviews and merges it — no human in the loop.
• The report is published as a shareable web page you can send to anyone.

Every step is a real, clickable link on the result page. You're not trusting a spinner — you're reading the agent's actual pull request. The whole loop takes about five minutes. If you're curious, the agents' work is all public: github.com/javimosch/cvboost-audits.

The three things it's really showing off

The audit tool is the demo. Here's the stack under it.

1. mago — agents that run companies

The audits aren't a script. They're done by mago, a platform for autonomous agents that take their work from a GitHub repo, ship pull requests, and escalate to a human only when they need to. A "company" is a repo; the agents are its team. For CVBoost, the team is an auditor that writes the report and a reviewer that merges it. They run 24/7 on your own machine with your own model key — you bring the keys, mago brings the orchestration. CVBoost is that loop running live, in the open, on every visitor's request.

2. hart — Claude Artifacts, self-hosted

Each report is a self-contained HTML page published with hart: one CLI call turns a finished HTML file into a live, sandboxed, versioned URL. It's the "publish this artifact" primitive that Claude has inside claude.ai — unbundled, open-source, and runnable on your own box, callable from any terminal agent. The auditor agent publishes your report to hart and hands you the link.

3. machin — one binary, no Node

The web app itself — the server-rendered marketing site, the reactive WebAssembly UI, and the JSON API — is a single static native binary written in machin, a language I've been building for the AI agent writing the code, not the human reading it. No Node, no bundler, no node_modules. One file you drop on a €3 VPS. The browser client is compiled to WebAssembly from the same language as the server, sharing the same model.

Why free, why public

CVBoost is free because it isn't the product — it's the shop window. The point is to let you watch an autonomous agent do real, useful work end to end, traceably, instead of reading a landing page that claims it can. If that makes you want agents doing this on your own repos, that's mago. If you want to publish HTML deliverables your agents generate, that's hart. And if you want to build the kind of tiny single-binary services that make this cheap to run, that's machin.

Go try it: audit your own profile, then open the pull request the agent left behind.