Skip to main content
PromptQuorum
Home/Power Local LLM/claude-code-local Review 2026: Run Claude Code on a Local Model
Coding Assistants

claude-code-local Review 2026: Run Claude Code on a Local Model

Β·12 min readΒ·By Hans Kuepper Β· Founder of PromptQuorum, multi-model AI dispatch tool Β· PromptQuorum

claude-code-local (github.com/nicedreamzapp/claude-code-local) is a free, open-source, unofficial, third-party project β€” it is not made by Anthropic and not officially affiliated with or endorsed by Anthropic. It runs a local MLX-native server on Apple Silicon that mimics the Anthropic API format Claude Code expects, so you can point Anthropic's official claude CLI at your own machine (via ANTHROPIC_BASE_URL) and have a different, locally-hosted open-weight model β€” such as Gemma, Qwen, Llama, or DeepSeek β€” answer instead of Anthropic's cloud-hosted Claude model. It does not run Claude itself offline; it repurposes the Claude Code client to talk to a local model through a compatible interface. The project is MIT-licensed, free, requires a Mac with Apple Silicon (M1 or newer), and is built and maintained by one independent developer, Matt Macosko, with no company or investor behind it.

claude-code-local (github.com/nicedreamzapp/claude-code-local) is a free, open-source, unofficial project that lets you keep using Anthropic's Claude Code command-line tool while a locally-run, open-weight model on your own Mac answers instead of Anthropic's cloud API. It is not made by Anthropic and not officially affiliated with or endorsed by Anthropic, and it does not run Anthropic's actual Claude model offline β€” it runs an MLX-native server that speaks the same API format Claude Code expects, so the CLI can be redirected to a different, locally-hosted model (Gemma, Qwen, Llama, DeepSeek, and others) via an API-compatible shim. This review covers what the project actually does, how to install it, its real hardware requirements, and where it fits next to other local coding-assistant setups.

Key Takeaways

  • Free and open source under the MIT License
  • Unofficial, third-party project β€” not made by, affiliated with, or endorsed by Anthropic, and it does not run Anthropic's actual Claude model
  • Runs an MLX-native, Anthropic-API-compatible server so the official claude CLI can be pointed at a local model via ANTHROPIC_BASE_URL
  • Requires a Mac with Apple Silicon (M1 or newer), Python 3.12, and Anthropic's Claude Code CLI installed separately
  • Supports several open-weight models sized to your Mac's unified memory, from Hermes 4 14B on 16 GB machines up to Qwen 3.5 122B on 96 GB+ machines
  • A v0.1.0 release (May 2026) fixed a bug where earlier setups leaked telemetry and background traffic to Anthropic's servers even with a local base URL configured β€” verify you are on a current version if privacy/air-gap use is your goal
  • Built and maintained by one independent developer, Matt Macosko, based in Arcata, California β€” no company or investor behind the project
  • GitHub repository (github.com/nicedreamzapp/claude-code-local) shows 3,314 stars and 625 forks as of September 2026

πŸ“ In One Sentence

claude-code-local is a free, open-source, unofficial MLX server that lets Anthropic's Claude Code CLI run against a locally-hosted open-weight model on Apple Silicon, instead of Anthropic's actual Claude model or cloud API.

πŸ’¬ In Plain Terms

It does not put Anthropic's Claude model on your computer. Instead, it makes your Mac pretend to be Anthropic's cloud so the claude command still works exactly as you know it, but a different, open-weight AI model that runs on your own hardware answers your questions β€” no account, no per-token bill, and no data sent to Anthropic's servers for those requests.

πŸ“ŒNote: This review is the deep-dive companion to claude-code-local's entry in the Local LLM Software Directory β€” see that page for how it compares at a glance to dozens of other local AI tools.

πŸ“ŒNote: Naming note: "claude-code-local" repurposes Anthropic's Claude Code client/CLI to talk to a different, locally-hosted open-weight model. It is not an Anthropic product, and it does not make Anthropic's Claude model run offline.

What Is claude-code-local?

claude-code-local is a local server, not a chat app or an IDE. It sits between Anthropic's official Claude Code CLI and a locally-run open-weight model, translating Claude Code's Anthropic-API-format requests into a format an MLX-hosted local model can answer, and translating the answer back. Its own GitHub description summarizes it as letting you "Run Claude Code 100% on-device with local AI on Apple Silicon" via an "MLX-native Anthropic-API server."

  • Core function: an Anthropic-API-compatible local server (proxy/server.py) that Claude Code can be pointed at instead of api.anthropic.com, via the ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY environment variables
  • Inference engine: MLX, Apple's array-computing framework for Apple Silicon, via the mlx-lm package β€” this is why the project is Mac-only and requires an M-series chip
  • What it is not: it does not download, run, or in any way expose Anthropic's actual Claude model. "Claude Code" in the project's name refers to Anthropic's CLI client, which this project repurposes to talk to a different, locally-hosted model
  • Developer: one independent developer, Matt Macosko, based in Arcata, California, per the GitHub README's footer β€” the README explicitly states "one person, no team, no investors"
  • Canonical repository: github.com/nicedreamzapp/claude-code-local β€” the original, active project; a same-named fork, jinzaizhichi/claude-code-local, exists on GitHub with zero stars and is not a separate independent project

Project History and Version Milestones

The GitHub repository was created on March 26, 2026. Development has been fast-paced but the project has kept its formal version tags to three releases so far, each bundling a significant fix or feature rather than shipping tags for every commit.

  1. 1
    v0.1.0 β€” May 8, 2026: Verified-offline fix and 16 GB Mac support
    Why it matters: A contributor found via `lsof` that Claude Code 2.1 was still reaching `api.anthropic.com` on startup for telemetry, feature flags, marketplace auto-install, and auto-updates, even with a local `ANTHROPIC_BASE_URL` set β€” meaning earlier setups following the old README were not actually fully offline. This release fixed it by disabling that background traffic via four documented environment variables, and switched the default model for 16 GB Macs (which had been crashing on the previous default) to a smaller, working model.
  2. 2
    v0.2.0 β€” August 5, 2026: Prompt-cache correctness fix
    Why it matters: Fixed a bug where the shared prompt cache became corrupted after the first session on a long-running server, per the release notes causing later sessions to answer from a previous, unrelated conversation's cached state. The project's own before/after measurement on one long-lived server went from 1 correct answer out of 8 to 8 out of 8 after the fix.
  3. 3
    v0.3.0 β€” August 22, 2026: Agent-12 benchmark launch and full-precision Qwen 3.8 27B support
    Why it matters: Introduced the developer's own open-source agent benchmark suite ("Agent-12") to help users pick a model, and added full-precision (non-quantized) support for a 27B model using speculative decoding for a claimed roughly 3-4x speedup. These performance and benchmark figures are the developer's own self-reported measurements β€” this review could not independently verify them, so treat specific numbers as a starting point rather than a guarantee for your own hardware.

What Can You Do With claude-code-local?

The project's feature set centers on making Claude Code work against a local model with minimal friction, plus several additional local-first modes built on the same underlying server. Here is what each part actually does, per the project's own GitHub README.

  • Claude Code, redirected locally β€” the core use case: keep using the claude CLI exactly as normal, but have it answer from a model running on your own Mac instead of Anthropic's cloud
  • "Keep going" fallback β€” a shell command that lets you continue an existing Claude Code conversation with a different model (a free OpenRouter cloud model, or a local one) when you hit Anthropic's own usage limit, without losing the conversation
  • Native Engine mode β€” a lighter-weight agent built into the project itself, offered as a faster alternative to routing every request through the full Claude Code client
  • Browser Agent mode β€” an optional mode where the locally-run model can drive a real browser window, documented separately in the project's browser-agent guide
  • Voice mode β€” an optional hands-free mode for talking to the local model and hearing a spoken response, documented in the project's voice-mode guide
  • Phone control β€” an optional mode for sending commands to the Mac-hosted server from a phone, documented in the project's phone-control guide
  • Traffic isolation β€” sets four of Claude Code's own documented environment variables to disable non-essential background network calls (telemetry, autoupdate, marketplace auto-install, background tasks), verifiable yourself via lsof -p $(pgrep -f claude)

Usage Examples: Three Ways to Use claude-code-local

These are concrete workflows built from the project's documented features above β€” not hypothetical use cases.

Pricing and Licensing: Is It Really Free?

Yes β€” claude-code-local itself is free and open source, under the MIT License, confirmed via the repository's license badge and LICENSE file. There is no subscription, account, or usage limit imposed by the project itself.

  • The server, installer, and launchers are free and MIT-licensed, with no paid tier
  • Local models run at no per-token cost, using your own Mac's hardware and electricity
  • The project still requires Anthropic's official Claude Code CLI, which is free to install; whether you separately maintain an Anthropic subscription is up to you and orthogonal to using this project
  • An optional "keep going" cloud fallback can use free-tier OpenRouter models with your own free API key β€” this is opt-in and not required for the core local workflow

Hardware Requirements and Model Choices

Hardware requirements scale directly with which model you choose to run, since claude-code-local has no fixed minimum of its own beyond Apple Silicon and macOS. The project's own setup script recommends a model tier based on your Mac's total unified memory.

Unified memory
Recommended model
16 GB (e.g. MacBook Air, base models)Hermes 4 14B
32–48 GB (Pro tier)Gemma 4 12B
64–95 GB (Max tier)Gemma 4 31B
96 GB and up (Max/Ultra tier)Qwen 3.5 122B, with room for Llama 3.3 70B and others alongside it

The project also documents several additional model options beyond the default per-tier recommendation, including Qwen 3.8 27B at full precision and a self-described "abliterated" (refusal-reduced) build of a 30B model called Muse-Glimmer. A note on abliterated models: the project's own README flags that these have their built-in safety refusals turned down and states plainly that "each model's own license still applies" and to "use them responsibly" β€” this is a real capability difference from a standard instruction-tuned model, and the responsibility for how you use one sits with you. Benchmark and speed figures for specific models throughout the README (including the Agent-12 leaderboard) are the developer's own self-reported measurements, not independently verified by PromptQuorum.

Who Should Use claude-code-local?

Whether this project fits depends on whether you specifically want to keep the Claude Code CLI workflow while swapping in a local model, and whether you have the Apple Silicon hardware it requires.

claude-code-local vs. Other Local Coding-Assistant Setups

claude-code-local is one of several ways to get AI coding assistance running against a local model instead of a cloud API. Here is how it sits next to other options in that space β€” see the Local LLM Software Directory for the full catalog.

  • oMLX β€” a general-purpose Apple Silicon MLX inference server with SSD-backed model caching; unlike claude-code-local, it is not built specifically around Claude Code's API format, but it is the closest comparison for anyone choosing an MLX-based local backend on a Mac. See the oMLX review.
  • Aider β€” a terminal-based, model-agnostic AI pair-programming tool that works with local models via any OpenAI-compatible endpoint, git-integrated by design; a good comparison if you want a coding CLI built local-first from the start rather than a shim in front of Anthropic's client. See the Aider review.
  • Cline β€” an open-source, VS Code-based autonomous coding agent that also supports local models through OpenAI-compatible endpoints; a good comparison if you prefer an IDE-embedded agent over a terminal-first workflow. See the Cline review.

Common Mistakes When Evaluating claude-code-local

Most confusion about this project comes from its name implying an official Anthropic offering, or from assuming any version is automatically fully offline.

Frequently Asked Questions

What is claude-code-local?

claude-code-local (github.com/nicedreamzapp/claude-code-local) is a free, open-source, unofficial MLX server for Apple Silicon Macs that lets Anthropic's Claude Code CLI connect to a locally-run open-weight model instead of Anthropic's cloud API.

Is claude-code-local made by Anthropic?

No. It is an independent, third-party, community-built project. It is not made by Anthropic and not officially affiliated with or endorsed by Anthropic in any way.

Does claude-code-local run Anthropic's actual Claude model offline?

No. It does not download, host, or run Anthropic's Claude model at all. It runs a different, locally-hosted open-weight model (such as Gemma, Qwen, Llama, or DeepSeek) and makes it answer through a server that speaks the same API format the Claude Code client expects.

Is claude-code-local free?

Yes. It is open source under the MIT License with no paid tier. You still need Anthropic's free Claude Code CLI installed as a dependency, but the project itself charges nothing.

What hardware does claude-code-local require?

A Mac with Apple Silicon (M1 or newer) and Python 3.12, since the project runs local inference through Apple's MLX framework. There is no Windows, Linux, or Intel Mac support. It works on a 16 GB Mac, per the project's own testing, with larger unified-memory Macs able to run larger, more capable models.

Does claude-code-local keep my code fully offline?

It can, but only from v0.1.0 onward. An earlier bug meant Claude Code itself reached Anthropic's servers for telemetry and background tasks even with a local base URL configured; v0.1.0 (May 2026) fixed this by disabling that traffic. Verify your own setup with lsof -p $(pgrep -f claude) rather than assuming.

How do I install claude-code-local?

The documented one-command installer is curl -fsSL https://raw.githubusercontent.com/nicedreamzapp/claude-code-local/main/install.sh | bash. A manual alternative β€” cloning the repository and running bash setup.sh after reading it β€” is also documented for anyone who prefers not to pipe a remote script into bash directly.

Which models can claude-code-local run?

The project documents several open-weight models sized by your Mac's unified memory, from Hermes 4 14B (16 GB) up to Qwen 3.5 122B (96 GB and up), plus additional options like Gemma 4, Llama 3.3 70B, and a full-precision Qwen 3.8 27B build. Model choice and quality vary by which one you pick and your hardware.

Who develops claude-code-local?

It is built and maintained by one independent developer, Matt Macosko, based in Arcata, California. The project's own README states it explicitly: "one person, no team, no investors."

When was claude-code-local first released?

The GitHub repository was created on March 26, 2026. The first formal version tag, v0.1.0, followed on May 8, 2026, with v0.2.0 (August 5, 2026) and v0.3.0 (August 22, 2026) since β€” see the official Releases page for anything shipped after this review's publish date.

Sources

← Back to Power Local LLM