The thesis
Most “free” AI tools subsidize inference with VC money or upsell cloud APIs. Neural Junkie inverts that:
- Free for users — MIT license, local-first runtime, optional cloud only if you add keys.
- Free to build — Tauri, React, Go, Ollama, GitHub Actions on a public repo.
- Free to market — static GitHub Pages site, GoatCounter, LinkedIn article pipeline, public benchmarks.
This page is the reproducible recipe. Clone the repo, read the tables, and run the same stack.
Build stack (OSS)
| Layer | Tool | License | Cost |
|---|---|---|---|
| Desktop shell | Tauri 1.x + Rust | MIT / Apache-2.0 | $0 |
| UI | React + TypeScript + Vite | MIT | $0 |
| Hub / agents | Go | BSD-3 | $0 |
| Local inference | Ollama (bundled macOS) | MIT | $0 |
| Editor | Monaco + LSP | MIT | $0 |
| Specialists | LoRA / MLX (Apple Silicon) | Apache-2.0 | $0 |
| CI | GitHub Actions | Free for public repos | $0 |
Ship & operate (free tiers)
| Function | Service | Cost |
|---|---|---|
| Source + issues | GitHub public repo | $0 |
| Installers | GitHub Releases | $0 |
| macOS install | Homebrew tap (camronwood/tap) | $0 |
| Marketing site | GitHub Pages (/docs) | $0 |
| Slack OAuth relay | Cloudflare Workers (optional) | Free tier |
| Analytics | GoatCounter | Free at our scale |
| SEO | Generated sitemap + meta (make site-seo-sync) | $0 |
Market (no ad spend)
- 47-page static site — landing, guides, download funnel, security, packs.
- Article pipeline —
docs/marketing/→make articles-sync→ long-form SEO pages. - Public benchmarks — local model pass rates as both engineering artifact and content.
- Gallery + release notes — screenshots and honest beta changelog.
- GitHub discoverability — topics, homepage, pack ecosystem repos.
User runtime: free path vs optional paid
| Need | Free path | Optional paid |
|---|---|---|
| Chat & agents | Ollama + bundled/local models | Claude, OpenAI-compatible APIs |
| Specialists | LoRA train locally | Cloud GPU (your choice) |
| Slack | Connect Slack via free relay | — |
| Storage | Local disk | — |
| Updates | GitHub Releases + in-app beta channel | — |
Platform taxes (honest limits)
Some gates cost money or user friction. We document both — no pretending Gatekeeper does not exist.
| Gate | Paid fix | Free path today |
|---|---|---|
| macOS Gatekeeper | Apple Developer + notarization ($99/yr) | Homebrew cask, Right-click → Open, install trust guide |
| Linux packages | Official apt repo / Snap store fees | Homebrew formula, .deb, build from source |
| Windows SmartScreen | EV code signing cert | “More info” → Run anyway, SHA256 verify, build from source |
| Official Homebrew Cask | Notability / maintainer time | Custom tap works today |
| Custom domain | Registrar fee | camronwood.github.io/neural-junkie (free) |
CI already supports notarization when Apple secrets are present — we stay on ad-hoc signing until paying Apple aligns with the experiment budget (zero).
Reproduce it yourself
git clone https://github.com/camronwood/neural-junkie.git
cd neural-junkie
make gui-install # first time
make gui # run desktop + hub
# Marketing site locally
python3 -m http.server 8765 --directory docs
# Regenerate SEO after editing pages
make site-seo-sync