Answer three questions. Watch an AI install your business.
FallForge is the concierge that installs FallHub for you. Point it at your existing website. It scrapes your footprint, distils it into a grounding kernel, packages your FallHub instance with the right vertical template and recommended modules, pre-configures the LLM router, and hands you a working AI-first business OS. Twenty minutes. Free. MIT.
What FallForge actually does.
Discover
Fetches your existing website and up to 5 signal pages (about · prices · reviews · contact · FAQ). Extracts text · links · social handles · listing URLs (Airbnb, Booking, Google). Detects your vertical with keyword fingerprinting.
Kernel distil
Runs the raw evidence through your chosen frontier LLM (default: Claude Opus 4.8). Extracts every fact into the vertical's kernel schema. Populates the not-present list to prevent hallucination. Learns your voice tone from real reply patterns.
Wire
Packages your bespoke FallHub install: kernel.json + BOM + install guide + module list. Downloads as a ZIP or (if you provide a GitHub token) pushes directly to a fresh repo. Configures the LLM router with your chosen blend baked in.
Calibrate
Post-install, your instance runs a training simulator (WebLLM Botler · free · local). Watches how you reply vs the AI's drafts. Auto-tunes confidence thresholds so your autopilot starts conservative and matches your voice.
Handoff
Live URL delivered. What's installed. Why. Cost projection at your chosen blend. Optional teach-me-the-router walkthrough. You slide the autonomy dial as trust grows.
Frontier at ramp · slide down as you trust it.
Every FallForge install ships with the FallRouter pre-configured. Pick your default blend at install time · change it any time via one dropdown. Your kernel doesn't move. Your tools don't rewire.
Full Frontier
Everything Claude Opus 4.8. Best quality. Highest cost.
80/20 (default)
Frontier for reasoning · Llama for bulk. Best ratio.
50/50 blend
Balanced tier. Cost-first with quality edges.
Full Open
Everything Llama 3.3 70B via Groq. Sovereign. Fast.
Local only
WebLLM Llama 3.1 8B. Browser only. Zero cost.
Sovereign by default. Hosted if you're in a hurry.
Local FallForge · free
Runs in your browser. Fetches your site directly. Distils via your chosen LLM (BYOK or free WebLLM). Nothing hits our servers. Slower on CORS-blocked hosts.
- Full sovereignty · your data stays yours
- No account required · nothing signed up for
- MIT · fork the FallForge repo, run it anywhere
- Uses your BYOK key or free WebLLM Botler
Hosted FallForge · £29 one-off
We run a Cloudflare Worker that handles the heavy work: CORS-free scraping, multi-page crawl, faster LLM calls, GitHub push assist. You still bring your own key · we broker the plumbing.
- 10-15% faster · handles more CORS-blocked hosts
- Deeper crawl (up to 50 signal pages)
- Automatic GitHub Pages deploy
- One-off charge · nothing recurring
Wishwood was forged this way.
Wishwood — the hospitality reference implementation for FallHub — was originally kernel-harvested through the same process. It's a real cottage, running the same substrate you'll deploy, taking real bookings today.
What you're wondering.
Does my data ever hit your servers?
Local mode: no. Everything runs in your browser. Your website scrape stays local. Your API key stays local. Hosted mode: yes — we run a Cloudflare Worker that helps with heavier fetching and GitHub push. You choose which mode. Sovereign is the default.
Do I need an AI subscription?
No. You can run the entire install on WebLLM (free · in-browser Llama 3.1 8B). It's slower but costs nothing forever. If you want faster or higher-quality distillation, plug in an API key from Anthropic, OpenAI, Gemini, Groq, or Mistral. Your key. Direct connection.
What happens after FallForge finishes?
You have a working FallHub instance. Landing page. Autopilot. Training simulator. Modules installed. Router configured. You open the training sim, practice with fake customers, then slide the autonomy dial as you trust it.
What if the vertical detection is wrong?
You can override at install time. The auto-detection is a keyword-fingerprint scan across your site text. If confidence is low, we ask before proceeding. If you're a bespoke niche not on the list, pick "closest fit" and edit the kernel afterwards.
What does the ZIP contain?
kernel.json (your distilled business facts) · bom.json (bill of materials) · README.md · INSTALL.md (step-by-step) · modules-to-install.txt. Drop the kernel into a FallHub fork, install the listed modules, you're live.