Brainrouter — AI skill for Claude Code

brainrouter

A Rust proxy and local control plane between coding harnesses and LLMs. Choose independent Main, Reviewer, and local Subagent roles, inspect current model activity and completed-stream measurements, and explore imported benchmarks without running benchmark jobs. Bonsai classification and Manifest cloud access are opt-in; the defaults stay local. Core management and reviews have a headless CLI, while the benchmark and model-observability pages also expose HTTP APIs.

coding harness -> brainrouter :9099 (OpenAI or Anthropic wire format)
                      |
                      +-- auto/default -> Main choice (local by default)
                      +-- local/cloud  -> explicit backend default
                      +-- model ID     -> exact local, or cloud/
                      +-- subs         -> separate local subagent pool
                      +-- review       -> snapshotted Reviewer choice
                      |
                      +-- local -> llama-swap
                      +-- cloud -> Manifest when enabled -> local fallback
                      +-- Auto role -> Bonsai when enabled, otherwise local

/dashboard  -> routing/profile/review/operations controls
/models     -> read-only activity, measured completions, indicative alerts
/benchmarks -> stored results, import previews, plan previews, inspector

• One endpoint, all harnesses. OpenAI-compatible on POST /v1/chat/completions. Anthropic-compatible on POST /v1/messages. Every harness connects to the same :9099.
• Independent routing roles. Choose Main, Reviewer (explicit local or cloud ID), and a separate local Subagent pool, or apply a named preset. Legacy auto, local, cloud, and subs aliases remain supported. Cloud and Bonsai stay off by default.
• Off by default, opt in. Fresh installs run fully local with a single hop — no Bonsai model downlo