# nextmillionai

> Open source, local-first alternative to YC's Paxel. It reads your AI coding sessions (Claude Code, Cursor, Codex) and turns them into a verifiable builder profile — entirely on your machine. "Runs locally" is enforced in CI, not just promised. Free forever, Apache-2.0.

nextmillionai measures *how* you build with AI — the part GitHub never saw. A commit stores the diff; the craft that produced it (how you prompt, steer, orchestrate agents, and recover) vanishes. nextmillionai keeps and scores it locally, then lets you share a portable, verifiable proof-of-work profile without your code ever leaving your machine.

## What it does today

- **Measures the "how"**: six dimensions (Signal Clarity, Build Stability, Decision Weight, Recovery Velocity, Context Command, Orchestration Range), scored by plain arithmetic over counted local signals against research-anchored bands. No model ever writes a score.
- **Counts the invisible work**: stats no other tool keeps — AI lines that survived, parallel agents, subagents dispatched, deep sessions, ship streaks.
- **Names your craft**: archetypes, work mode, and a build-domain map — a map of how you work, never a ranking.
- **Is verifiable**: every number is cross-corroborated across session transcripts, git history, and a local ledger, stamped with sources and a confidence score.
- **Is shareable on your terms**: publish a derived-only profile (no code, no prompts), revocable any time.

## Why it's different

- **Built for a new craft**: commits, résumés, and interviews were built for a world before agents; this is the first instrument made to measure building with AI.
- **A map, not a ladder**: no percentiles, no leaderboards, no "top 10%". Every kind of builder is a craft, scored for fit — never ranked against another person.
- **Open by construction**: exactly one file in the codebase is allowed to touch the network, and CI fails the build if any other file tries. Open source (Apache-2.0), runs on your machine.

## The Paxel context

YC launched Paxel on June 6, 2026 with the promise "it runs locally inside Docker, and your code never leaves your machine." Within hours the community found it uploading prompt excerpts, file paths, and git metadata to YC servers, and the launch posts were taken down. nextmillionai is the version where that is structurally impossible: open source, local-first, and CI-enforced.

## Install

```
git clone https://github.com/nextmillionai/nextmillionai && cd nextmillionai
python3 -m nextmillionai start
```

## Links

- [GitHub repository](https://github.com/nextmillionai/nextmillionai): source, scoring methodology, and the CI-enforced network boundary
- [network.py](https://github.com/nextmillionai/nextmillionai/blob/main/nextmillionai/network.py): the only module allowed outbound network access
- [CI network rule](https://github.com/nextmillionai/nextmillionai/blob/main/tests/test_network.py): the test that fails the build if any other file touches the network
- [Documentation](https://nextmillionai.org/docs): how to install and use the tool
- [Scoring methodology](https://nextmillionai.org/methodology): how every number is computed
- [nextmillionai.org](https://nextmillionai.org): the main project site

## Facts

- License: Apache-2.0 (open source)
- Price: free forever for developers
- Privacy: assessment runs offline; nothing uploaded unless you explicitly publish (derived signals only, revocable)
- Supported tools: Claude Code, Cursor, Codex, plus git
- Install: `git clone https://github.com/nextmillionai/nextmillionai && cd nextmillionai && python3 -m nextmillionai start`