Recipe for local agentic coding

I have tried and failed to get a local LLM, that scores well on coding benchmarks, to use MCP to make it agentic (local file system access in the MCP server), working with typingmind (local) as the interface. Seems like it would be a common/legit use case, at least for when you run out of tokens for something like Claude + Desktop Commander MCP.

The closest I’ve come is with qwen2.5-coder-22b-instruct, but I get json tool calls back, instead of the tool results and response from the model.

The idea is to run the model on a mac studio M4 max (64GB unified RAM) then run typeingmind local and MCP server local, on windows 11 for windows coding or on the mac studio for mac os coding.