Sous — AI skill for Claude Code
Run Claude Code's subagents on a local LLM (MLX, Qwen) on Apple silicon and stretch your Pro/Max usage limits: a local daemon that serves subagent turns from the local model and proxies everything els.
How to install Sous
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open krcm0209/sous and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Sous does
Run Claude Code's subagents on a local LLM (MLX, Qwen) on Apple silicon and stretch your Pro/Max usage limits: a local daemon that serves subagent turns from the local model and proxies everything else to Anthropic unchanged. Hybrid local + cloud — Claude keeps the main loop, the local model does the subagent work, Claude Code runs every tool.
Alternatives in AI
- UI UX Pro Max Skill — An AI SKILL that provide design intelligence for building professional UI/UX multiple platforms 45.8k ★
- Vllm Mlx — High-performance OpenAI and Anthropic compatible LLM inference server for Apple Silicon 1.5k ★
- Agmsg — Cross-vendor messaging for CLI AI coding agents — let Claude Code, Codex, Gemini & Copilot talk to each other 1.5k ★
README
sous
Keep Claude on Claude Code's main loop and run its subagents on your Mac.
sous is a local daemon for Apple silicon that sits between Claude Code and Anthropic. Claude Code's main conversation — the planning, the decisions, the review — goes to Anthropic exactly as before, on your own Pro/Max login. The subagents it spawns (Explore, general-purpose, your own agents) are answered by a local MLX model instead, so the tokens they generate don't count against your plan. Claude Code still runs every tool itself, under its own permission mode; sous runs nothing. Claude is the head chef; the local model does the prep.
┌─► api.anthropic.com the main loop and every other request,
Claude Code ──► sous daemon ──┤ forwarded byte for byte on your login
127.0.0.1 └─► local MLX model subagent turns (model id sous-local)
You need an Apple silicon Mac with ~18 GB of unified memory free for the default model (a ~16 GB download), and a Claude Pro or Max plan. Only sessions you start with `sous claude` go through sous; plain `claude` is untouched.
Why use it
Heavy Claude Code use runs into plan usage limits, and a large share of what spends them can be output a subagent generates: files read and summarised, code searched, a module drafted. sous moves that output onto your Mac, where a token costs no plan usage, and leaves Claude the part that needs a frontier model: deciding what to delegate and reviewing what comes back.
- Claude stays in charge. The full-local setups below point every Claude Code request, the main loop included, at the local model. sous forwards the main loop untouched, so only delegated work drops to local-model quality.
- One install, one launcher. sous is both the router and the model
server.
sous claudestarts Claude Code wired to it: no proxy to pair with a server, no API key, no settings file edited, and your login is never stored or re-issu
Related Skills
Ovo Local LLM
A private Claude-Code-style coding agent for Apple Silicon — run chat, code, and local model workflows on-devi
Setup Lmstudio VSCode
Set up a fully local coding model in VS Code on an Apple Silicon Mac (LM Studio MLX engine, no admin), with au
Claude Failover
Backup brain for Claude Code. When your Claude Max plan hits its limit or Claude is down, flip one command and
Omp Subrouter
Route omp (oh-my-pi) through a local manaflow-ai/subrouter daemon — pooled Claude Max / ChatGPT subscriptions
Mac Code
mac code — Claude Code, but it runs on your Mac for free. 35B AI agent at 30 tok/s via Apple Silicon flash-pag
Wax
Shared Single-file memory layer for all your agents, sub mili-second RAG over text, photo and video on Apple S
Related Agents
Buildhost Lead
Use PROACTIVELY and automatically — do not wait to be asked — as the Claude-Code node lead FOR an Apple-silico
Cost Estimator
Estimates Cloudflare Workers costs before deploying. Reads wrangler.toml, counts D1 tables/rows, estimates KV/
Rs React Code Quality Check
Subagent of rs-school-react-pr-reviewer. Inspects everything except TypeScript and lint/format/bundler — namel