MIT licensed · open source

Let your coding agent offload the grunt work.

A tiny MCP server that hands well-specified sub-tasks to a local model running in Ollama — on your own GPU — instead of spending cloud tokens on boilerplate.

View on GitHub Setup guide

How it fits together

your coding agent (e.g. Claude Code) │ │ MCP tool call: ask_local_model ▼ offload — this repo, one file │ │ HTTP ▼ Ollama — running on your machine │ ▼ a local model, on your GPU

The agent decides what to delegate and reviews what comes back. This is just the wire — not a framework, not an opinion about when to use it.

Two tools, nothing else

`ask_local_model`

Sends a prompt (with an optional system prompt and model override) to your local Ollama instance and returns the response.

`list_local_models`

Lists what you've actually pulled, so the agent can check before assuming a model is available.

Quick start

ollama pull qwen2.5-coder:7b-instruct

git clone https://github.com/ToxinNozaki/offload.git
cd offload && npm install

Then point your MCP client at index.js. Full instructions, including the .mcp.json snippet for Claude Code, are in the README.

What this isn't

Not a replacement for your main coding agent — it's a tool that agent calls.
Not an orchestration framework. No task queues, no review gates, no multi-agent pipeline — build that on top if you want it.
Not tied to Claude Code. Anything that speaks MCP over stdio can use it.