Claude Code Local: Run the Full Agent Loop on Your MacBook With Zero Cloud Calls

A GitHub repo called claude-code-local has quietly crossed 2,500 stars by solving a specific problem: running Claude Code's agent workflow entirely on your laptop, with no API keys and no data leaving your machine. The project spins up a lightweight local server that speaks the same protocol Claude Code expects, but routes every request to an on-device model instead of Anthropic's API.

What Claude Code Local Actually Does

The repo, built by Matt Macosko, stands up a proxy server on your Mac that intercepts Claude Code's outbound calls. Instead of hitting Anthropic's cloud endpoint, the proxy redirects those calls to a local model running through tools like Ollama or llama.cpp. The result is the same agentic coding workflow (file reads, edits, terminal commands, multi-step reasoning) powered by open-weight models running on your own hardware.

The README lists support for three models across four modes: Qwen 3.5 (the default at 17.6 billion parameters), Llama 3.3 70B, and Gemma 4 31B. The video specifically highlights Qwen 3.5 at 102 billion parameters running at 65 tokens per second, which suggests users with enough VRAM can push beyond the defaults. The project is MIT-licensed with over 500 forks.

The Privacy Angle: AirGap AI

The most compelling demo shown in the video is what the repo calls AirGap AI. The screen recording shows a 70-billion-parameter Llama model auditing a confidential legal document (a non-disclosure agreement) with Wi-Fi physically turned off and lsof running to prove zero outbound network connections. The overlay text reads "0 OUTBOUND" while the model processes the NDA on-device.

This matters for anyone working with sensitive codebases, legal documents, or proprietary business logic. Running the full Claude Code tool loop locally means your prompts, file contents, and model outputs never touch a third-party server.

Who This Is For

Claude Code Local is practical for a few specific use cases:

Developers handling sensitive code who want the agentic workflow without sending proprietary source code to an API
Teams under compliance requirements (HIPAA, SOC 2, NDA-bound work) where data cannot leave the local environment
Hobbyists and tinkerers who want to experiment with Claude Code's agent loop using free, open-weight models
Cost-conscious developers who want to avoid per-token API charges for routine coding tasks

The tradeoff is model quality. Open-weight models at 17B to 102B parameters will produce noticeably different results than Anthropic's frontier models. But for many coding tasks (refactoring, file manipulation, boilerplate generation), the local models may be good enough.

Key Takeaways

claude-code-local is an open-source proxy that reroutes Claude Code's API calls to a model running on your MacBook
It supports Qwen 3.5, Llama 3.3, and Gemma 4 out of the box, with four operating modes
The AirGap AI demo proves zero network activity while processing confidential documents locally
The project has 2,600+ stars, 500+ forks, and an MIT license
You get the same agent loop (file edits, terminal, multi-step reasoning) without API keys or cloud dependency

Resources

claude-code-local on GitHub (search for "claude-code-local" by zapp)
Ollama for running local models on macOS
Qwen 3.5 model page on Hugging Face