> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/janhq/jan/llms.txt
> Use this file to discover all available pages before exploring further.

# jan launch

> Start a local model and launch an AI agent with it pre-wired

## Usage

```bash theme={null}
jan launch [PROGRAM] [OPTIONS] [-- PROGRAM_ARGS...]
```

Start a local model, then launch an AI agent or program with environment variables automatically set to use the local model.

## Arguments

<ParamField path="program" type="string">
  Agent or program to run after the model is ready.

  Supported agents:

  * `claude` — Claude Code (Anthropic's AI coding agent)
  * `codex` — Codex CLI (OpenAI's coding agent)
  * `openclaw` — OpenClaw (open-source autonomous AI agent)

  Omit to pick interactively from the list above.

  ```bash theme={null}
  jan launch claude
  jan launch openclaw
  jan launch  # Pick interactively
  ```
</ParamField>

<ParamField path="program_args" type="array">
  Arguments forwarded to the program. Use `--` to separate them from `jan` options.

  ```bash theme={null}
  jan launch claude -- --help
  jan launch openclaw -- tui --debug
  ```
</ParamField>

## Options

### Model Selection

<ParamField query="model" type="string">
  Model ID to load. Omit to pick interactively from installed models.

  ```bash theme={null}
  jan launch claude --model qwen3.5-35b-a3b
  ```

  Can be:

  * A model ID from `jan models list` (e.g. `qwen3.5-35b-a3b`)
  * A HuggingFace repo ID (e.g. `Qwen/Qwen2.5-35B-Instruct-GGUF`) — will auto-download
</ParamField>

<ParamField query="bin" type="string">
  Path to the inference binary. Auto-discovered from Jan data folder when omitted.

  ```bash theme={null}
  jan launch claude --model qwen3.5-35b --bin /usr/local/bin/llama-server
  ```
</ParamField>

### Server Configuration

<ParamField query="port" type="number" default="6767">
  Port the model server listens on.

  ```bash theme={null}
  jan launch claude --model qwen3.5-35b --port 8080
  ```
</ParamField>

<ParamField query="api-key" type="string" default="jan">
  API key for the model server. Exported as `OPENAI_API_KEY` and `ANTHROPIC_AUTH_TOKEN`.

  ```bash theme={null}
  jan launch claude --model qwen3.5-35b --api-key my-secret-key
  ```
</ParamField>

### Performance Configuration

<ParamField query="n-gpu-layers" type="number" default="-1">
  GPU layers to offload.

  * `-1`: All layers (full GPU acceleration)
  * `0`: CPU only
  * `> 0`: Specific number of layers to offload

  ```bash theme={null}
  jan launch claude --model qwen3.5-35b --n-gpu-layers 35
  ```
</ParamField>

<ParamField query="ctx-size" type="number" default="4096">
  Context window size in tokens.

  ```bash theme={null}
  jan launch claude --model qwen3.5-35b --ctx-size 8192
  ```

  <Note>
    Setting `--ctx-size` explicitly disables auto-fit. Omit this option when launching Claude to enable auto-fit by default.
  </Note>
</ParamField>

<ParamField query="fit" type="boolean">
  Auto-fit context to available VRAM.

  ```bash theme={null}
  jan launch claude --model qwen3.5-35b --fit
  ```

  **Default behavior**:

  * When launching `claude`: `--fit` defaults to `true` (unless `--ctx-size` is set explicitly)
  * When launching other programs: `--fit` defaults to `false`
</ParamField>

### Output Control

<ParamField query="verbose" type="boolean" default="false">
  Print full server logs (llama.cpp / mlx output) instead of the loading spinner.

  ```bash theme={null}
  jan launch claude --model qwen3.5-35b --verbose
  jan launch claude --model qwen3.5-35b -v
  ```
</ParamField>

## Examples

<CodeGroup>
  ```bash Basic usage theme={null}
  # Pick agent and model interactively
  jan launch

  # Launch Claude Code with a specific model
  jan launch claude --model qwen3.5-35b-a3b

  # Launch OpenClaw
  jan launch openclaw --model qwen3.5-35b-a3b
  ```

  ```bash With arguments theme={null}
  # Pass arguments to the launched program
  jan launch claude -- --help
  jan launch openclaw -- tui --debug
  ```

  ```bash Custom configuration theme={null}
  # Custom port and API key
  jan launch claude --model qwen3.5-35b --port 8080 --api-key secret

  # Maximize context window
  jan launch claude --model qwen3.5-35b --fit

  # CPU only
  jan launch claude --model qwen3.5-35b --n-gpu-layers 0
  ```

  ```bash HuggingFace auto-download theme={null}
  # Download and launch with a HuggingFace model
  jan launch claude --model Qwen/Qwen2.5-35B-Instruct-GGUF
  ```
</CodeGroup>

## Environment Variables

Depending on the launched agent, Jan sets different environment variables:

### For Claude Code & Codex

```bash theme={null}
OPENAI_BASE_URL=http://127.0.0.1:6767/v1
OPENAI_API_KEY=jan
OPENAI_MODEL=qwen3.5-35b-a3b
ANTHROPIC_BASE_URL=http://127.0.0.1:6767
ANTHROPIC_AUTH_TOKEN=jan  # or ANTHROPIC_API_KEY for non-Claude agents
ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3.5-35b-a3b
ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3.5-35b-a3b
ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3.5-35b-a3b
```

### For OpenClaw

Jan configures OpenClaw via `~/.openclaw/openclaw.json` instead of environment variables:

```json theme={null}
{
  "models": {
    "providers": {
      "jan": {
        "baseUrl": "http://127.0.0.1:6767/v1",
        "apiKey": "jan",
        "api": "openai-completions",
        "models": [
          {
            "id": "qwen3.5-35b-a3b",
            "name": "qwen3.5-35b-a3b",
            "input": ["text"],
            "reasoning": false,
            "contextWindow": 131072,
            "maxTokens": 16384,
            "cost": {
              "input": 0,
              "output": 0,
              "cacheRead": 0,
              "cacheWrite": 0
            }
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "jan/qwen3.5-35b-a3b"
      }
    }
  }
}
```

Provider API keys are cleared to prevent overrides:

* `OPENAI_API_KEY`, `OPENAI_BASE_URL`
* `ANTHROPIC_API_KEY`, `ANTHROPIC_AUTH_TOKEN`, `ANTHROPIC_OAUTH_TOKEN`
* `GEMINI_API_KEY`, `MISTRAL_API_KEY`, `GROQ_API_KEY`
* `XAI_API_KEY`, `OPENROUTER_API_KEY`

## Output

### Loading

```
✓ qwen3.5-35b-a3b ready · http://127.0.0.1:6767

  OPENAI_BASE_URL=http://127.0.0.1:6767/v1
  OPENAI_API_KEY=jan
  OPENAI_MODEL=qwen3.5-35b-a3b
  ANTHROPIC_BASE_URL=http://127.0.0.1:6767
  ANTHROPIC_AUTH_TOKEN=jan
  ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3.5-35b-a3b
  ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3.5-35b-a3b
  ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3.5-35b-a3b

  → Launching: claude
```

Then Claude Code (or your chosen agent) starts with the local model pre-configured.

### For OpenClaw

```
✓ qwen3.5-35b-a3b ready · http://127.0.0.1:6767

  ~/.openclaw/openclaw.json → jan provider configured
  agents.defaults.model.primary = jan/qwen3.5-35b-a3b

  → Launching: openclaw tui
```

## Workflow

When you run `jan launch`:

1. **Model Selection**: Pick a model (interactively or via `--model`)
2. **Model Loading**: Start the model server with optimized settings
3. **Environment Setup**:
   * For Claude/Codex: Set environment variables
   * For OpenClaw: Write `~/.openclaw/openclaw.json`
4. **Agent Launch**: Execute the chosen program with the local model pre-wired
5. **Cleanup**: When the agent exits, stop the model server

## Supported Agents

### Claude Code

```bash theme={null}
jan launch claude --model qwen3.5-35b-a3b
```

Launches Anthropic's Claude Code agent with:

* `ANTHROPIC_AUTH_TOKEN` set to your API key (default: `jan`)
* `ANTHROPIC_BASE_URL` pointing to local model
* All Anthropic model defaults (`OPUS`, `SONNET`, `HAIKU`) set to your local model
* Auto-fit enabled by default for maximum context window

### Codex CLI

```bash theme={null}
jan launch codex --model qwen3.5-35b-a3b
```

Launches OpenAI's Codex CLI with:

* `OPENAI_API_KEY` set to your API key (default: `jan`)
* `OPENAI_BASE_URL` pointing to local model
* `OPENAI_MODEL` set to your local model ID

### OpenClaw

```bash theme={null}
jan launch openclaw --model qwen3.5-35b-a3b
```

Launches the open-source OpenClaw agent with:

* `~/.openclaw/openclaw.json` configured with Jan provider
* Default model set to `jan/<model_id>`
* OpenClaw's TUI automatically started (unless you pass custom args)

## Performance Tips

### Maximize Context for Claude Code

```bash theme={null}
jan launch claude --model qwen3.5-35b
```

Auto-fit is enabled by default when launching Claude, maximizing the context window based on available VRAM.

### Custom Context Size

```bash theme={null}
jan launch claude --model qwen3.5-35b --ctx-size 8192
```

Setting `--ctx-size` explicitly disables auto-fit and uses your specified size.

### CPU-Only Mode

```bash theme={null}
jan launch claude --model qwen3.5-35b --n-gpu-layers 0
```

## HuggingFace Auto-Download

Download and launch with a HuggingFace model:

```bash theme={null}
jan launch claude --model Qwen/Qwen2.5-35B-Instruct-GGUF
```

Jan will:

1. Fetch available GGUF files
2. Let you pick a quantization
3. Download the model
4. Start the model server
5. Launch Claude Code

For private/gated models:

```bash theme={null}
export HF_TOKEN="your_token_here"
jan launch claude --model meta-llama/Llama-3.3-70B-Instruct-GGUF
```

## Troubleshooting

### Agent Not Found

```
Error launching 'claude': No such file or directory
```

**Solution**: Ensure the agent is installed and available in your `PATH`.

```bash theme={null}
# For Claude Code
which claude

# For OpenClaw
which openclaw
```

### Model Not Found

```
Error: model not found in data folder
```

**Solution**: Download the model using the Jan desktop app, or use a HuggingFace repo ID to auto-download.

### Binary Not Found

```
✗ llama-server binary not found
```

**Solution**: Install a backend from Jan's settings, or specify the binary path with `--bin`.

### OpenClaw Config Override

If OpenClaw still uses a cloud provider instead of Jan:

**Solution**: Check for environment variables that override the config:

```bash theme={null}
env | grep -E '(OPENAI|ANTHROPIC|GEMINI|MISTRAL|GROQ|XAI|OPENROUTER)'
```

Jan clears these automatically, but they may be set in your shell config (`.bashrc`, `.zshrc`, etc.).

## See Also

<CardGroup cols={2}>
  <Card title="Serve Command" icon="server" href="/cli/serve">
    Run models without launching an agent
  </Card>

  <Card title="Commands Reference" icon="terminal" href="/cli/commands">
    Complete reference for all CLI commands
  </Card>
</CardGroup>