LLM tooling · for Claude Code
Make Claude Code punch above its quota.
brainmux builds tools that route your work to the right model — so Claude Code spends your Opus quota on architecture and review, not busywork.
What is it
brainmux is LLM tooling for Claude Code. llmproxy routes Claude Code to cheap OpenRouter models through local LiteLLM proxies and delegates the grunt work to them — one key reaches thousands of models, the brains run pay-as-you-go and never touch your Anthropic quota, and Opus stays the orchestrator. graphmux gives those agents a local, deterministic code graph, so they ground on your codebase instead of guessing. Each tool wraps a mature open-source core (LiteLLM, CodeGraph) in a thin, pinned layer — same core, our packaging.
Products
Two tools today. A family in the making.
llmproxy
Live · v0.1Run Claude Code on cheap LLM brains and delegate the grunt work — one OpenRouter key, thousands of models, your Opus quota untouched.
Explore llmproxy ↓graphmux
Live · v0.1Give Claude Code and bmux delegates a local, deterministic code graph — real callers, callees and impact — so agents ground on your codebase instead of guessing.
Explore graphmux ↓llmproxy · our first tool
Run Claude Code on cheap brains.
Your Opus quota is the bottleneck. llmproxy routes Claude Code to cheap OpenRouter models and delegates the grunt work to them — Opus stays the orchestrator; cheap brains do the volume.
Thousands of models
A single OpenRouter key reaches DeepSeek, Qwen, GLM, GPT, Gemini and hundreds more — no per-provider setup.
Never touches your quota
The brains run pay-as-you-go on OpenRouter. Your Anthropic subscription quota stays untouched.
Opus stays in charge
Cheap brains do the volume — bulk edits, detection sweeps. Opus reviews, decides, and fixes.
How it works
One config. One proxy per brain. Routed by port.
A single brains.yaml (zod-validated) is the source of truth. bmux generates a Docker Compose stack — one LiteLLM proxy per brain, isolated by port — and points Claude Code at the brain you choose.
What you get
A CLI for the whole loop.
Manage the stack
init · up / down · health · config add-brain / set-model · test. Declarative — edit, regenerate, restart.
Offload grunt work
bmux delegate coder "…" hands a bounded task to a cheap brain headless. Opus verifies the result — no rubber-stamp.
Pick from the live catalog
Browse the live OpenRouter catalog by price, context and use-case — never a guessed, stale model slug.
Spend & logs, per brain
Each brain ships the LiteLLM UI for spend, request logs and parameter tuning. Nothing to rebuild.
Quickstart
Four commands to a running brain.
Requires Docker and an OpenRouter API key.
graphmux · our second tool
Ground agents on your codebase.
Cheap brains hallucinate about code. graphmux gives them a local, deterministic code graph — real callers, callees, impact and verbatim source — so they look up your codebase instead of guessing. A thin wrapper over the vendored CodeGraph engine (tree-sitter → local SQLite): no embeddings, no cloud, telemetry off.
Real call-graph, not guesses
tree-sitter-grade callers, callees and blast-radius — exact references, not approximate vector matches.
Your code never leaves
Indexes to a local SQLite graph. No embeddings API, no cloud, and the vendored engine's telemetry is forced off.
Cheap brains stop inventing
bmux delegate --memory wires the graph into a cheap brain, so it queries real symbols before it acts.
FAQ
Questions, answered.
What is brainmux?
brainmux is LLM tooling for Claude Code. Its first tool, llmproxy, lets you run Claude Code on cheap alternate LLM models and delegate grunt work to them — so your Opus subscription quota goes to architecture and review, not busywork.
What is llmproxy?
llmproxy is a Claude Code plugin (CLI: bmux) that routes Claude Code to cheap OpenRouter models via local LiteLLM proxies. One OpenRouter key reaches thousands of models across providers like DeepSeek, Qwen, GLM, GPT and Gemini.
What is graphmux?
graphmux is the second brainmux plugin (CLI: gmux). It gives Claude Code and bmux delegates a local, deterministic code graph — real callers, callees, impact and verbatim source — so agents ground on your actual codebase instead of guessing. It's a thin wrapper over the vendored CodeGraph engine (tree-sitter to a local SQLite graph): no embeddings, no cloud.
Does graphmux send my code anywhere?
No. graphmux indexes your repo into a local SQLite graph on your own machine — no embeddings API, no cloud upload, and the vendored engine's telemetry is forced off by default. Install it with `/plugin install graphmux@brainmux`, then `gmux install` and `gmux index`.
How does graphmux stop cheap brains from hallucinating?
Run `bmux delegate <brain> --memory "<task>"`. It wires graphmux's code-graph MCP into the cheap brain (isolated — no host MCP noise), so the brain looks up real callers, callees and impact before it acts, instead of inventing file paths or symbol names.
Does it use my Anthropic or Opus quota?
No. The cheap brains run pay-as-you-go on OpenRouter, on a separate meter that never touches your Anthropic subscription quota. Opus stays the orchestrator; the cheap brains do the volume.
What models can I use?
Any model on OpenRouter — DeepSeek, Qwen, GLM, Kimi, GPT, Gemini and hundreds more — with a single key. `bmux models` browses the live catalog by price, context and use-case, so you never guess a stale model slug.
How much does it cost?
llmproxy is free and MIT-licensed. You only pay OpenRouter's pay-as-you-go per-token price for the models you actually use — usually cents. Your Anthropic subscription quota is untouched.
How do I install it?
In Claude Code, run `/plugin marketplace add brainmuxhq/brainmux` then `/plugin install llmproxy@brainmux`. Then `bmux init`, add your OpenRouter key (`bmux config add-key OPENROUTER_API_KEY`), `bmux up`, and `bmux test`.
Do I need Docker?
Yes. Each brain runs as a local LiteLLM Docker container isolated by port, with one shared Postgres. brainmux generates the whole Docker Compose stack from a single brains.yaml file.