Your AI keeps working when the API goes dark.
FallRouter is a sovereign kernel that routes every LLM request between your local Ollama models and frontier APIs (Claude · OpenAI · Gemini). When a provider hits rate limits, gets export-controlled, goes bankrupt, or has a three-week outage — your app doesn't. The local leg picks up automatically. Zero downtime. Zero dependency. Tamper-evident audit chain. TBAnth replaces per-token API bills.
The sales pitch
Three weeks. That's how long the frontier went dark for.
You know it. Your clients know it. When frontier API access disappears — export controls, rate limits, provider outages, price changes — everything downstream stops. Every business that built on hosted AI just watched its product go dark.
"Our AI didn't go down."
That's the sentence your clients want to be able to say. FallRouter makes it true. Every prompt has a local fallback. Every routing decision is audited. Every failure reroutes automatically. Users see slightly different model output — never a broken screen. This is what "sovereign AI" actually looks like at runtime, not in marketing decks.
Cost reduction
Local leg handles ~90% of tokens · frontier only for the hard 10%
Failover time
Health-check every 30s + retry with smallest local on any error
Vendor lock-in
MIT-licensed source · swap Ollama, Claude, OpenAI in one config line
Audit trail
SHA-256 prevHash chain in SQLite · one command verifies
How the routing works
Femto · Nano · Milli · Advisor.
Every request gets classified in <1ms by a deterministic sieve (length, code content, keywords). Then the right leg fires. No ML in the classifier. No vendor lock-in. Predictable behaviour.
Femto<50 chars · route/classify/yes-no
Fires the smallest local model (3B). Instant. Free. Handles the routing decisions themselves, plus quick user acks.
Nano50-500 chars · write/explain
Local 8B model. Fast. Sovereign. The workhorse for typical chat / drafting / explanation tasks. Runs on any modern GPU or capable CPU.
Milli>500 chars or code · deep reasoning
Routes to frontier if available (best-in-class reasoning). Falls back to local 70B if the API is unavailable. Zero drop-off.
Advisor"check"/"verify"/"audit" keywords · high-stakes
Fires the specific frontier model chosen as the advisor (usually Claude Opus). The 10% of calls that earn the spend. Auto-fallback on unavailability.
Auto-failover · what happens when the frontier goes dark
- Router pings each configured provider every 30s (cheap HEAD requests · <3s timeout)
- On any request, checks last availability snapshot before routing
- If preferred leg is dark → routes to next available (frontier → other frontier → local)
- If mid-request the preferred leg errors → retries once against smallest local (the "ALWAYS WORKS" fallback)
- Every reroute logged to the audit chain (tier, model, reason, latency, cost)
- Your app sees a valid response · never sees a 503
Access
Three tiers. Start free.
Sovereign tier is free forever · you download it, you run it, you own it. The paid tiers add convenience (managed configs, hosted dashboard, memory layer, priority support).
Sovereign
- OpenAI-compatible router service
- Femto/nano/milli/advisor classifier
- Auto-failover to local
- Single-HTML sovereign dashboard
- SQLite tamper-evident audit chain
- Cost tracker · routing log
- Docker one-command deploy
- MIT source · fork · modify · redistribute
Hybrid
- Everything in Sovereign
- Managed frontier API failover pool
- Real-time reroute notifications
- Cost anomaly alerts
- Config profiles per app
- Email support · 48h response
Client
- Everything in Hybrid
- Vector-DB persistent memory (episodic + semantic)
- Multi-model routing (per-conversation persona)
- White-label dashboard for your clients
- Custom classifier tuning
- Priority support · 24h response
- Included setup call (60 min)
Quick start
Sovereign tier running in 5 minutes.
Two prerequisites: Node 20+ and Ollama installed with at least one model pulled.
1 · Get the source
Clone the repo (or download the release tarball). Everything is in-tree — router service, dashboard, SDK adapter, Docker stack, docs.
2 · Start Ollama + pull models
The router expects the classifier model, the standard model, and (optionally) the deep model. Small enough to run on any modern GPU or capable CPU.
3 · Run the router service
Set the master key in .env. Add any frontier API keys you want the auto-failover to consider. Then npm start.
4 · Open the dashboard
The single-HTML dashboard connects to the router service. Enter the URL + master key in Settings (stored in your browser's IndexedDB · nothing leaves this device).
5 · Send a test request
Send anything from the Test tab. Watch the routing log fill in real-time. Kill your internet — watch the auto-fallback kick in.
FAQ
The questions people actually ask.
http://localhost:4100/v1 with your master key. Zero code changes. The router transparently picks whether the actual call goes to Ollama, Anthropic, or OpenAI based on the classifier.