Documentation · AI & cost
AI models & routing
The platform's work is performed by named agents — implementation, architecture, devops, business understanding, and more. Each agent is routed to a model of your choice, with an optional fallback. You decide what intelligence runs where, and on whose bill. For the end-to-end setup steps behind each option below — who's responsible for what, and how to validate it's working — see Platform, tenant, or your own subscription.
Your model options
| Option | How it works | Billing |
|---|---|---|
| Platform models | Curated models operated by the platform — zero setup. | Metered against your prepaid balance, per token. |
| Your own API keys (BYOK) | Add your Anthropic/OpenAI-compatible provider credential; calls go straight to your provider account. | Your provider bills you; the platform only meters usage for visibility. |
| Self-hosted models | Point an Ollama custom provider at a machine you run — good for cheap auxiliary tasks. | Free (your hardware). |
| Your Claude subscription | Claude Code on your build VM, dev-mode — full guide. | $0 (subscription limits apply). |
Routing, fallbacks, and resilience
- Primary + fallback per agent: if the primary model fails or is rate-limited, the call retries on the fallback automatically. Pair a subscription/self-hosted primary with a paid fallback for uninterrupted delivery.
- Circuit breaker: a model failing repeatedly is bypassed in favor of its fallback for a cooldown window instead of burning full timeouts on every call.
- Spend controls: set a monthly budget cap (BYOK) — calls are refused, with a clear error, once it's reached. Platform-model usage stops when prepaid balance runs out. Auto-deliver pauses itself on budget blocks rather than failing silently.
- Batch discount: background documentation work can run on the 50%-discounted batch tier — a toggle in Admin, on by default, off if you want everything synchronous.
Efficiency features like prompt caching and surgical repair edits are always on — they're engineering, not configuration. The ledger shows their effect per call (cache read/write columns in Recent invocations).