04 / brouter

Route the task. Not the request.

An independent AI-native task router for your real configured agents. Your Markdown formulas decide where work goes. Not an LLM request router — bifrost handles that layer.

A CLI of its own.

Use bro router inside the toolkit, or install @broject/brouteron its own. Standalone brouter does not require the bro CLI. Connect the agents you use, define the routing policy, and send each task to the right one.

npm i -g @broject/bro
bro router --help

# Or use brouter independently
npm i @broject/brouter
npx brouter --help

Route to your configured agents, not imaginary models. Keep your clients, credentials and working sessions. brouter makes the routing decision; the selected agent does the work.

Inspect the package →

You write the formula. Code keeps the rails.

User-authored routing formulas in Markdown describe the questions and scoring. The router routes tasks to the agents you configured — not imaginary models. Code owns health, chains and hard filters; the judge answers typed questions. Cache off by default. Unknown cost stays unknown.

task→MD formula→configured agent→outcome

Outcome feedback closes the routing loop. Learn from completed work, compare formula choices against real results, and carry that evidence into the next decision.

Prove the lift.

One prompt, bare agent versus agent + bro. Measure the workflow's contribution separately from router replay quality. Compare outcomes, cost and latency on the same task, so you can see what the agent does alone and what bro adds.

Read the routing model →