Hybrid routing & budgets.

How routing works

Rabbit uses a capability-aware routing policy: tool-calling and agentic work is only ever sent to models that can handle it, and cheaper steps prefer local or mid-tier models. You steer it with a small set of settings and let the fallback chain handle the rest.

Settings

Routing is driven by Rabbit AI settings (and the headers they produce):

rabbit-ai.routing.execution       auto     # auto · cloud · local  (X-Rabbit-Execution)
rabbit-ai.routing.preferProvider  auto     # or pin: anthropic · openai … (X-Rabbit-Prefer-Provider)
rabbit-ai.budget.dailyUsd         5.00     # soft daily cap — warns, does not block
rabbit-ai.budget.warnAtUsd        4.00     # threshold for the spend warning

execution expresses your preference for where a turn runs;preferProvider pins a provider slug when you want one, and falls back automatically when it is unavailable or incapable.

Fallback chains

Each capable route resolves through a fallback chain -- local → mid-tier cloud → frontier. When the primary model is unavailable (rate-limited, network error, budget warning), the router walks the chain until a model responds. A local fallback keeps the agent from stalling.

Budgets

rabbit-ai.budget.dailyUsd is a soft daily cap: once your spend approaches it, Rabbit surfaces a warning in the chat and status bar -- it does not hard-stop a conversation. Hard caps apply to background agent spawns, so a runaway fleet can't blow past your limit.

Cost tracking

Token usage and estimated cost are shown in the status bar in real time, and the session cost chip tracks per-conversation spend as you go.