The routine work an agent gets handed — a summary, an extraction, a commit message, a first look at a diff — runs on your own hardware with a prompt already tuned for it, which keeps both the budget and the paid model's context for the parts that need them.
A set of purpose-built tools that route work to your local Ollama instance. Each ships with a baked-in system prompt and a description written for the calling agent, so the handoff happens at the right moment and comes back tuned rather than raw.
- A log, a document or a transcript summarised locally, with a summariser prompt already in place
- A first-pass code review from the local model, behind a reviewer prompt
- A commit message drafted from the change
- Structured values pulled out of text — every URL in a file, and the like — behind an extractor prompt
- A plain question or an ongoing chat sent to the local model when the answer does not need frontier reasoning
Ollama running locally with at least one model pulled — the project's example is qwen2.5-coder:14b. Python 3.11+, or uvx which manages it for you. Two dependencies, mcp and httpx, and no telemetry. Everything runs on your machine, so the cost is your GPU rather than tokens.
One command — uvx ollama-handoff