Apr 17 2026 How to Run Claude Code with a Local LLM on Apple Silicon

Configure Claude Code to use a local model served by LM Studio on Apple Silicon, with a practical setup based on LM Studio's Anthropic-compatible API.

This guide shows you how to run Claude Code with a locally hosted large language model on Apple Silicon. The supported integration path is to let LM Studio expose an Anthropic-compatible API endpoint and then point Claude Code at that local endpoint via environment variables.

The examples below use LM Studio on localhost:1234 and a locally loaded model with a 32K context window.

Prerequisites

A Mac with Apple Silicon.
LM Studio installed.
claude installed.
A local model downloaded or imported into LM Studio.
Enough available RAM for the model you want to run.

Install Claude Code

Install Claude Code using one of the current supported methods.

Native install:

curl -fsSL https://claude.ai/install.sh | bash

Or via Homebrew:

brew install --cask claude-code

After installation, verify that the CLI is available:

claude --version
claude doctor

Part 1: Prepare LM Studio

Start LM Studio

Open LM Studio at least once. Then verify that the lms CLI is available in your shell.

Check that lms is installed:

lms --help

If this command is not found, follow the LM Studio CLI setup/bootstrap step from the LM Studio documentation and then reopen your shell.

List locally available models:

lms ls

If your model files were downloaded outside LM Studio, you can import them:

lms import /path/to/model.gguf

Load a Model with a 32K Context Window

Load your chosen model into memory and set the context length explicitly.

If you want to estimate memory usage before loading:

lms load --estimate-only <model_key> --context-length 32768

lms load <model_key> --context-length 32768

You can also assign a stable identifier for API use:

lms load <model_key> --context-length 32768 --identifier qwen-local

To see which models are currently loaded:

lms ps

Start the Local Server

Start LM Studio’s local server:

lms server start --port 1234

Check server status:

lms server status

At this point, LM Studio serves an Anthropic-compatible Messages API on http://localhost:1234/v1/messages.

Part 2: Point Claude Code at LM Studio

Configure Environment Variables

Set Claude Code to use your local LM Studio server:

export ANTHROPIC_BASE_URL=http://localhost:1234
export ANTHROPIC_AUTH_TOKEN=lmstudio

If LM Studio’s Require Authentication option is enabled, replace lmstudio with your LM Studio API token.

When using ANTHROPIC_BASE_URL plus ANTHROPIC_AUTH_TOKEN against LM Studio, Claude Code authenticates against the local endpoint and does not need the usual browser login flow for that session.

Choose the Model

You can select the model when launching Claude Code:

claude --model qwen-local

If you did not assign a custom identifier, use the loaded model name that LM Studio exposes.

Alternatively, set the model via an environment variable:

export ANTHROPIC_MODEL=qwen-local
claude

Part 3: Test the Local Endpoint

Before starting Claude Code, test LM Studio directly against the Anthropic-compatible endpoint.

curl http://localhost:1234/v1/messages \
  -H "content-type: application/json" \
  -H "x-api-key: lmstudio" \
  -d '{
    "model": "qwen-local",
    "max_tokens": 128,
    "messages": [
      {
        "role": "user",
        "content": "Write a one-line hello message."
      }
    ]
  }'

If this request succeeds, Claude Code should be able to use the same local server.

Part 4: Run Claude Code in a Project

Start Claude Code inside your project directory:

cd /path/to/your/project
claude --model qwen-local

Typical workflow:

Ask Claude Code to inspect files in the repository.
Let it propose or apply edits.
Run your local build or test commands.
Review diffs before committing changes.

For example, in a Pelican project:

pelican content

And then commit as usual:

git add .
git commit -m "Update article"

Troubleshooting

Claude Code cannot connect

Check that the LM Studio server is running:

lms server status

Check that the model is loaded:

lms ps

Check that the environment variables are set in the same shell where you launch claude:

echo "$ANTHROPIC_BASE_URL"
echo "$ANTHROPIC_AUTH_TOKEN"

Responses are slow

Agentic coding workloads are context-heavy. A local model may perform noticeably better with a larger context window. A practical starting point is around 25K tokens or more, and 32K is a reasonable target if your hardware can support it.

If performance is poor:

Try a smaller or more coding-oriented model.
Reduce other RAM-heavy workloads.
Lower the context length if the model does not fit comfortably.
Test a shorter prompt first to verify baseline responsiveness.

Need server or prompt diagnostics

Stream LM Studio logs:

lms log stream --source server

Stream model input and output:

lms log stream --source model --filter input,output

Include prediction stats when available:

lms log stream --source model --filter output --stats

Notes

This setup uses LM Studio’s Anthropic-compatible /v1/messages endpoint. It does not require a custom JSON provider file, and it does not use the old /v1/completions integration pattern.

If you want a persistent setup, you can place the same environment variables in your shell profile, for example ~/.zshrc:

export ANTHROPIC_BASE_URL=http://localhost:1234
export ANTHROPIC_AUTH_TOKEN=lmstudio
export ANTHROPIC_MODEL=qwen-local

Summary

With LM Studio serving a local model through its Anthropic-compatible Messages API and Claude Code pointed at that local endpoint, you can run a local-inference coding workflow on Apple Silicon.

In short, the essential steps are:

Load a model in LM Studio with your desired context length.
Start the LM Studio local server.
Set ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN.
Launch Claude Code with --model or ANTHROPIC_MODEL.