Bitrouter

BitRouter

Help the user understand and reduce their LLM spend by routing their agent's model traffic through BitRouter. This project is analyzed in place; nothing is uploaded, and the user's own bitrouter.yaml is never overwritten.

When to use

• The user asks how much they could save on LLM/model costs.
• The user wants a routing policy, or to tighten an existing one.
• The user asks what models this codebase calls and where cheaper routing fits.

Tools

• estimate — Scan the codebase locally and estimate the savings from routing through BitRouter. Writes .bitrouter/analysis.json. No policy file.
• audit — Estimate plus a written review of routing opportunities. Writes bitrouter-audit.md and .bitrouter/analysis.json. No policy file.
• optimize — Tighten the routing policy from real metered traffic in the local bitrouter.db. Writes an observed report and a proposed bitrouter.optimized.yaml. Degrades to a cold-start audit when there is not enough traffic yet.

All three default to an offline / observed-offline path — deterministic and with no network egress. Each returns a short summary and the artifact paths.

Command

• /bitrouter — Run the observed optimizer over the current project and report the observed savings headline.

Live widget

While a session is active, a widget above the editor shows the observed savings headline when there is metered traffic, or a cold-start hint when there is not.

Invariants

• Zero-write to your inputs. The tools never modify your environment, code, or existing bitrouter.yaml. Proposals go to bitrouter.proposed.yaml or bitrouter.optimized.yaml.
• Local first. Estimate/audit scan locally; optimize reads a copy of your metering DB. Nothing is uploaded.