Connect Claude Code to LM Studio: local LLM setup on Apple Silicon

Prerequisites

• A Mac with Apple Silicon and enough unified memory for the model you want (see "Choosing a model" below; 32 GB is comfortable, 16 GB works with small models and a shorter context).
• LM Studio 0.4.1 or newer, opened at least once.
• Claude Code: curl -fsSL https://claude.ai/install.sh | bash or brew install --cask claude-code, then claude --version.

Step 1: Load a model in LM Studio

LM Studio ships a command-line tool called lms. If your shell does not find it, run the bootstrap step from the LM Studio app (Developer tab) and reopen the terminal.

lms get qwen/qwen3-coder-30b   # download from LM Studio's catalog (see note below)
lms ls                         # models on disk
lms load --estimate-only qwen/qwen3-coder-30b --context-length 32768
lms load qwen/qwen3-coder-30b --context-length 32768 --identifier qwen-local
lms ps                         # what is loaded, with its context length

Two settings matter more than anything else here:

Context length is the first of them: Claude Code sends a system prompt and tool definitions of roughly 20K tokens before your first word. LM Studio's own guidance is "more than ~25k", while the model's default in LM Studio is often 4K, and with that the session fails on the first request. Use 32K as the floor and 64K if memory allows, because the KV cache grows with it.

Tool calling is the second: the agent edits files and runs commands through tool calls. A model without reliable tool use will chat about your code and never touch it. Coding models with tool support that run well on Apple Silicon:

• qwen/qwen3-coder-30b (30B mixture-of-experts, 3B active; about 19 GB at 4-bit) — the best balance I have found for agentic work; wants 32 GB.
• openai/gpt-oss-20b (about 12 GB loaded with a 32K context; LM Studio's and Ollama's default suggestion)—the choice for 24 GB machines.
• qwen/qwen3.8-27b (dense 27B, about 18 GB)—the current general model in LM Studio's catalog, strong at tool use; needs 32 GB or more once the context is loaded.
• google/gemma-4-12b (about 7 GB on disk, 12 GB loaded with 32K)—the practical option for a 16 GB Mac; it completed file-editing tasks reliably in my tests.

Where the model comes from matters, so download through LM Studio's catalog (lms get <model> or the app's Discover tab), which carries chat templates LM Studio has tested. A GGUF taken from an arbitrary Hugging Face repository can have a template that LM Studio's Jinja engine cannot render once an agent sends its tool definitions, and the plain curl test below then works while Claude Code fails with a 500 (see troubleshooting).

Memory on a 16 GB Mac is the limiting factor: the KV cache for a 32K context adds several gigabytes, so a 7 GB model needs about 12 GB loaded, and LM Studio's resource guardrail refuses loads it expects to freeze the machine. Use --context-length 16384 or a smaller model there, and close memory-hungry applications.

Step 2: Start the server and check it with curl

lms server start --port 1234
lms server status

The Anthropic-compatible endpoint of LM Studio is http://localhost:1234/v1/messages. Check it before involving Claude Code—first the model list, then one message:

curl http://localhost:1234/v1/models      # "qwen-local" must be in the list
curl http://localhost:1234/v1/messages \
  -H "content-type: application/json" \
  -H "x-api-key: lmstudio" \
  -d '{"model":"qwen-local","max_tokens":64,
       "messages":[{"role":"user","content":"Say hello in one line."}]}'

A JSON reply with "type":"message" means the server side is done. Note that /v1/models lists every downloaded model when LM Studio's just-in-time loading is on (the default), whether it is loaded or not, so a model name in that list does not mean the model is in memory. lms ps shows what is loaded and with which context, and curl http://localhost:1234/api/v0/models reports a state of loaded or not-loaded per model. The custom identifier (qwen-local) exists only while the model is loaded.

Step 3: Point Claude Code at LM Studio

export ANTHROPIC_BASE_URL=http://localhost:1234
export ANTHROPIC_AUTH_TOKEN=lmstudio
export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen-local
export ANTHROPIC_DEFAULT_SONNET_MODEL=qwen-local
export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen-local
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=32768
claude --model qwen-local

The lines in that snippet do the following:

• ANTHROPIC_BASE_URL sends every request to LM Studio.
• ANTHROPIC_AUTH_TOKEN is the credential. LM Studio ignores its value unless Require Authentication is on, in which case use your LM Studio API token. Setting it also tells Claude Code not to use your claude.ai login for the session.
• The three ANTHROPIC_DEFAULT_*_MODEL lines map Claude Code's model aliases to the local model. This is the step most guides leave out: Claude Code runs some background work on its haiku alias, and if that alias still resolves to an Anthropic model name, LM Studio answers 404 and the session breaks in odd places. With the mapping in place, /model sonnet and friends all mean "the local model".
• CLAUDE_CODE_MAX_CONTEXT_TOKENS tells Claude Code how large the loaded context is. For a model name it does not recognise it assumes 200K tokens and prints [claude-code:unrecognized_model] at startup; with the real value it compacts the conversation before LM Studio rejects it. Match the --context-length you loaded.
• --model qwen-local selects the model for this session; ANTHROPIC_MODEL does the same from the environment.

Older guides mention ANTHROPIC_SMALL_FAST_MODEL for the background model, which is deprecated in favour of ANTHROPIC_DEFAULT_HAIKU_MODEL.

For a fully offline session add DISABLE_TELEMETRY=1, DISABLE_ERROR_REPORTING=1 and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1. Claude Code reads only whether these are set, so any value works. LM Studio's integration page also suggests the variable CLAUDE_CODE_ATTRIBUTION_HEADER=0.

Inside Claude Code, /status shows which base URL and credential source are in use—the first thing to check when something is off.

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:1234",
    "ANTHROPIC_AUTH_TOKEN": "lmstudio",
    "ANTHROPIC_MODEL": "qwen-local",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "qwen-local",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "qwen-local",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "qwen-local",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "32768"
  }
}

Step 4: Work

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

Ask it to inspect the repository, propose edits, run the tests, and review the diff before you commit. Keep tasks small and concrete, because local models do far better with "add a --dry-run flag to this script" than with "clean up the project".

Troubleshooting

Connection refused

Nothing listens on the port. lms server status; start the server; confirm the port in ANTHROPIC_BASE_URL.

Claude Code shows the login screen although curl works

The credential variable is not set in the shell you launched claude from, or a settings file overrides it. echo $ANTHROPIC_AUTH_TOKEN; check /status.

A startup warning about two credential sources ("auth may not work as expected")

Both ANTHROPIC_API_KEY and ANTHROPIC_AUTH_TOKEN are set. Keep one; LM Studio accepts either header.

404 or "model not found" in the middle of a session

A background call used the haiku alias. Set ANTHROPIC_DEFAULT_HAIKU_MODEL (and the sonnet/opus ones) to your local model as above.

500 Error rendering prompt with jinja template

Followed by "Unknown StringValue filter: safe" or "Cannot apply filter trim to type: Undefined". The model file's chat template does not survive the tool definitions Claude Code sends; the simple curl works because it sends none. I hit this with a community GGUF of Qwen3-Coder. Download the model again from LM Studio's catalog (lms get qwen/qwen3-coder-30b), which uses a corrected template, or override the template under My Models → Prompt Template. Nothing on the Claude Code side is wrong.

[claude-code:unrecognized_model] at startup, or "assumes a context window of 200k"

Informational: Claude Code does not know the local model's window. Set CLAUDE_CODE_MAX_CONTEXT_TOKENS to the loaded context length.

"claude.ai connectors are disabled because ANTHROPIC_API_KEY or another auth source is set"

Expected when ANTHROPIC_AUTH_TOKEN is set: the session uses the local credential and skips the claude.ai features that need your login. Harmless.

400 naming a token or context limit

The loaded context is too small. Reload with --context-length 32768 or more; lms ps shows the value in force.

curl http://localhost:1234/v1/models lists the model, Claude Code still fails

The model is downloaded but not loaded (lms ps says so). With just-in-time loading on, the first request loads it with the model's default settings, usually a 4K context, and the session fails with the 400 above; with just-in-time loading off, the request gets a 404. Load the model yourself with lms load and --context-length before starting Claude Code, as in step 1.

The model explains what it would do but never edits a file

No (working) tool calling. Switch to one of the models listed above.

The first answer takes a minute, later ones are fast

Prompt processing of the ~20K-token system prompt. Normal; larger context and a bigger model make it longer.

Everything is slow

The model does not fit comfortably. Use a smaller quantisation or model, lower the context length, close memory-hungry apps. lms log stream --source server and lms log stream --source model --filter output --stats show what the server is doing.

Summary

1. lms load <model> --context-length 32768 --identifier qwen-local
2. lms server start --port 1234
3. export ANTHROPIC_BASE_URL=http://localhost:1234 ANTHROPIC_AUTH_TOKEN=lmstudio, map the ANTHROPIC_DEFAULT_*_MODEL aliases to the local model and set CLAUDE_CODE_MAX_CONTEXT_TOKENS to the loaded context
4. claude --model qwen-local

Related guides: Claude Code with Ollama on macOS and Linux and OpenAI Codex CLI with a local model.

Sources: LM Studio, Claude Code integration · LM Studio, Anthropic compatibility endpoints · Claude Code, connect to an LLM gateway · Claude Code, model configuration.