---
title: "Open-Source claude-code-local Runs Anthropic's Coding Agent 7.5x Faster Offline"
subtitle: "claude-code-local runs Anthropic's coding agent on Apple Silicon with zero API fees, zero cloud calls, and 65 tok/s using open-source models"
url: "https://alphasignal.ai/news/open-source-claude-code-local-runs-anthropic-s-coding-agent-7-5x-faster-offline"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-06-18T21:25:47.000Z"
updated: "2026-06-30T19:12:32.352Z"
type: "Repo"
topics:
  - "Development"
  - "Gpus"
subtopics:
  - "Code Generation"
  - "Edge Deployment"
  - "Quantization"
upvotes: 19
source_url: "https://github.com/nicedreamzapp/claude-code-local"
comments: 0
truncated: true
---

# Open-Source claude-code-local Runs Anthropic's Coding Agent 7.5x Faster Offline

> claude-code-local runs Anthropic's coding agent on Apple Silicon with zero API fees, zero cloud calls, and 65 tok/s using open-source models

[claude-code-local](https://github.com/nicedreamzapp/claude-code-local) is an open-source project that lets you run Anthropic's Claude Code agent entirely on your Mac , no API key, no cloud, no monthly bill. It works by spinning up a local server that speaks the Anthropic API natively, so Claude Code thinks it's talking to Anthropic's cloud while the actual inference runs on your Apple Silicon GPU. The repo has crossed 2.8k stars and 540 forks since launch, and the stack has grown well beyond its original single-model design.

## The core problem it solves

Claude Code speaks the Anthropic API. Local models speak the OpenAI API. Different languages , so everyone builds a proxy to translate between them, and that proxy adds latency, complexity, and breaks things. The standard approach chains three processes together: Claude Code, a translation proxy, and the model backend. claude-code-local eliminates the middle layer entirely.

The result is dramatic. TurboQuant compresses the KV cache 4.9x, but the proxy was the bottleneck , killing it and speaking the Anthropic API directly made things 7.5x faster. A coding task that took 133 seconds with Ollama or llama.cpp plus a proxy now completes in 17.6 seconds.

## Pick your fighter

The project runs Claude Code 100% on-device with an MLX-native Anthropic-API server, supporting Qwen 3.5 122B at 65 tok/s, Llama 3.3 70B, and Gemma 4 31B , private, offline, and airgap-ready, built for NDA, legal, and healthcare workflows. Each model targets a different hardware tier:

* **Gemma 4 31B** , \~15 tok/s, \~18 GB RAM, fits a 32 GB Mac. Best for daily coding.
* **Qwen 3.5 122B** , 65 tok/s, \~75 GB RAM, needs a 96 GB+ Mac. Qwen wins raw speed because it's an MoE , only 10B of 122B params activate per token.
* **DeepSeek V4 Flash** , 284B total / 37B active, \~32 tok/s, 1M token context, requires 128 GB. It ships with on-disk KV cache so a 25k-token Claude Code system prompt prefills exactly once, ever.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [nicedreamzapp/claude-code-local — GitHub](https://github.com/nicedreamzapp/claude-code-local)
- [Google's TurboQuant: The Unsexy AI Breakthrough Worth Watching — Stark Insider](https://www.starkinsider.com/2026/03/google-turboquant-llm-compression-less-memory.html)
- [audiohacking/claude-code-local — GitHub (fork)](https://github.com/audiohacking/claude-code-local)
- [Run Claude Code Locally on Apple Silicon Using LM Studio and LiteLLM — Medium](https://medium.com/data-science-collective/run-claude-code-locally-on-apple-silicon-using-lm-studio-and-litellm-zero-cost-1416a6b984af)