Free offline AI coding agent. No subscription. Works on phones, RPi, anything.
npm install # One dependency: express
node setup.js # Download model (~4GB, one-time)
node core.js # Start server
# Open http://localhost:8899| Tier | What | Setup |
|---|---|---|
| 1: PWA | Browser + home screen install | Open http://<server>:8899 on phone |
| 2: Capacitor | Native Android/iOS app shell | cd mobile/capacitor && npm install && node build-mobile.js |
| 3: WebLLM | On-device AI in browser (WebGPU) | Tap 🧠 On-Device in app header |
- Run
node core.json your laptop/RPi - Open
http://<server-ip>:8899/connect.htmlon any device - Scan the QR code with your phone
cd mobile/capacitor
npm install
node build-mobile.js --server http://192.168.1.10:8899
npm run add:android # first time
npm run android # open Android StudioSee mobile/capacitor/README.md.
Requires Chrome 127+ with WebGPU and 8GB+ RAM. Enable via the header toggle. See mobile/webllm/README.md.
| Mode | What it does |
|---|---|
| Server mode | Uses llama-server (fastest, persistent) |
| CLI mode | Uses llama-cli (works everywhere) |
| Learning mode | No model needed — explains concepts |
| On-device (WebLLM) | Model runs in browser via WebGPU |
- 🧠 ARC Cycle — Analyzes → Plans → Codes → Self-Reviews
- 💡 Idea Mode — Describe your idea, get questions, then code
- 📱 Phone Ready — PWA, voice input, share sheet, wake lock
- 🔌 Fully Offline — After model download, zero internet
- 💾 Projects — Save and load your work
- 🖥️ CLI — Terminal mode for headless systems
node deploy.js test # Verify readiness
node deploy.js ship east-africa # Full kit + docs + training + partners
node deploy.js kit south-asia # Community Kit only
./deploy.sh all global # Shell wrapperRegions: east-africa, south-asia, southeast-asia, latin-america, west-africa, global
Output goes to deploy/ — Community Kits, offline docs, i18n, Navigator training, partner packages, impact dashboard.
CI/CD (ships all 13 formats): push a version tag and GitHub Actions builds everything:
git tag v1.0.2 && git push origin v1.0.2Manual:
node release.js --dry-run
node release.js allSee RELEASE_CHECKLIST.md.
- Node.js 18+
- 4GB free disk space (server model)
- 2GB+ RAM (4GB recommended; 8GB+ for on-device WebLLM)
This repo includes the Cursor Gate stack for agent-driven code review. Gate reviewer deps live in requirements-gate.txt (Node runtime deps remain in package.json).
pip install -r requirements-gate.txt
bash scripts/install-agent-environment.sh
bash scripts/gate-file.sh --file samples/hello_passing.pyPR CI: workflow Gate Check runs the gate on samples/hello_passing.py. Gap audit: docs/ROADMAP.md.
Gate scope note: core.js and deploy.js are not individually gated yet — they are large runtime modules. CI gates samples/hello_passing.py as the agent-policy smoke test. Critical JS changes should be reviewed manually or via future sample wrappers (see ROADMAP milestone 2).
Frontier integration (PR #3) was reverted on main; the path forward is the mirrored project-nexus/ tree, which tracks zowskyy/project-nexus until the Cursor GitHub App has push access.
cd project-nexus
python3 build/arc_orchestrator.py --slides 15See project-nexus/MIRROR.md and project-nexus/README.md for architecture. Lighthouse runtime (core.js) remains Node/Express; Frontier re-integration is planned once the standalone project-nexus repo is gate-ready.
MIT — Free forever.