Three months after Claude Code became the default coding companion for a growing number of developers, one of its users noticed a pattern: the tool was burning tokens on tasks like file searches and grep commands that a local model could handle for free.
Locode is an open-source CLI built around that observation. It tries to run simple, mechanical tasks (finding files, searching codebases, basic edits) on a local LLM running on your own machine, and only routes complex reasoning and multi-step problem solving to Claude's API.
The concept is sensible. If you have watched Claude Code work through a task, you have probably seen it spend several API calls on operations that feel like they should not require a frontier model. A local model running on even modest hardware can handle "find all TypeScript files in src/" without sending anything to Anthropic's servers.
The project is early and experimental, available on GitHub. The hard part will be getting the routing right - deciding which tasks are "simple enough" for a local model and which genuinely need Claude's reasoning. Get that boundary wrong in either direction and you either waste money or get bad local results that need to be redone anyway.