# Jan CLI (/desktop/cli) Available since **Jan 0.7.8** (March 2026). The `jan` CLI lets you serve local AI models and launch autonomous agents from your terminal — no cloud account, no usage fees, full privacy. ``` ██╗ █████╗ ███╗ ██╗ ██║██╔══██╗████╗ ██║ ██║███████║██╔██╗██║ ██ ██║██╔══██║██║╚████║ ╚█████╔╝██║ ██║██║ ╚███║ ╚════╝ ╚═╝ ╚═╝╚═╝ ╚══╝ Jan runs local AI models (LlamaCPP / MLX) and exposes them via an OpenAI-compatible API, then wires AI coding agent like Claude Code directly to your own hardware — no cloud account, no usage fees, full privacy. Models downloaded in the Jan desktop app are automatically available here. Usage: jan Commands: serve Load a local model and expose it at localhost:6767/v1 (auto-detects LlamaCPP or MLX) launch Start a local model, then launch an AI agent with it pre-wired (env vars set automatically) threads List and inspect conversation threads saved by the Jan app models List and load models installed in the Jan data folder help Print this message or the help of the given subcommand(s) Options: -h, --help Print help (see a summary with '-h') -V, --version Print version Examples: jan launch claude # pick a model, then run Claude Code against it jan launch claude --model janhq/Jan-code-4b-gguf # use a specific model jan serve janhq/Jan-code-4b-gguf # expose a model at localhost:6767/v1 jan serve janhq/Jan-code-4b-gguf --fit # auto-fit context to available VRAM jan serve janhq/Jan-code-4b-gguf --detach # run in the background jan models list # show all installed models ``` Models downloaded in the Jan desktop app are automatically available to the CLI. ## Installation [#installation] Jan CLI is installed automatically when you launch the Jan desktop app for the first time — no extra steps needed. You can uninstall or reinstall it at any time from **Settings > General > Jan CLI**. On macOS/Linux the installer writes the binary to `/usr/local/bin/jan` when that directory is writable, otherwise it falls back to `~/.local/bin/jan`. Make sure the chosen directory is on your `$PATH` so the `jan` command works from any terminal. The Jan desktop app helps you manage inference backends (LlamaCPP, MLX) and models — or use the CLI to manage them. Both share the same data folder, so models installed via either interface are available to both. ## Quick Start [#quick-start] Getting started takes a single command: ```bash jan launch ``` Jan will ask you to pick an agent (Claude Code or OpenClaw), then automatically download and set up Jan's foundation model and wire it to the agent for you. No config files, no API keys, no cloud — your agent runs entirely on your own hardware. --- ## Commands [#commands] ### `jan serve` [#jan-serve] Load a local model and expose it at `localhost:6767/v1` as an OpenAI-compatible API. Auto-detects LlamaCPP or MLX. ```bash jan serve [MODEL_ID] [OPTIONS] ``` | Option | Description | Default | |--------|-------------|---------| | `MODEL_ID` | Model ID to load (omit to pick interactively). Can be a local model ID or a HuggingFace repo ID (e.g., `unsloth/Qwen3.5-9B-GGUF`) | — | | `--port` | Port to listen on (0 = random free port) | `6767` | | `--api-key` | API key required by clients | `""` | | `-d, --detach` | Run in background, print PID | — | | `--embedding` | Treat model as an embedding model (not yet supported with llamacpp on the CLI — load embedding models from the desktop app) | — | | `--timeout` | Seconds to wait for the model server to become ready | `120` | | `-v, --verbose` | Print full server logs | — | Since Jan 0.8.0, llama-server runs in **router mode** under a single process. The flags `--ctx-size`, `--n-gpu-layers`, `--threads`, and `--fit` are accepted by the CLI for backwards compatibility but are currently **ignored** — per-model inference settings come from the router preset (`/llamacpp/router.preset.ini`) generated by the desktop app. Tune them from the desktop UI for now. When no model ID is provided, an interactive selector is shown. If no models are installed yet, Jan will automatically download its default foundation model to get you started: ``` $ jan serve ━━━ Select Model ━━━ Choose a model: > janhq/Jan-v3-4B-base-instruct-gguf [LlamaCPP] sentence-transformer-mini [LlamaCPP] Jan-v3-4B-base-instruct-4bit [MLX] ``` **Examples:** ```bash jan serve # pick a model interactively jan serve qwen3.5-35b-a3b # serve a specific model jan serve qwen3.5-35b-a3b --fit # auto-fit context to available VRAM jan serve qwen3.5-35b-a3b --detach # run in background jan serve qwen3.5-35b-a3b --port 8080 # serve on a custom port jan serve unsloth/Qwen3.5-9B-GGUF # download and serve a HuggingFace model ``` When using a HuggingFace repo ID (e.g., `unsloth/Qwen3.5-9B-GGUF`), if the model isn't downloaded yet, Jan will automatically download it from HuggingFace. #### LlamaCPP Server Arguments [#llamacpp-server-arguments] Environment variables with the `LLAMA_ARG_` prefix are forwarded to the underlying llama-server router process at startup. Per-model defaults (context size, GPU layers, threads, flash-attn, KV cache type, etc.) come from the router preset generated by the desktop app at `/llamacpp/router.preset.ini` — edit a model's settings in the desktop UI to change them. ```bash export LLAMA_ARG_HOST=0.0.0.0 # bind the router to a specific host jan serve qwen3.5-35b-a3b ``` --- ### `jan launch` [#jan-launch] Start a local model, then launch an AI agent with it pre-wired — environment variables are set automatically so the agent connects to your local model. ```bash jan launch [PROGRAM] [OPTIONS] ``` | Option | Description | Default | |--------|-------------|---------| | `PROGRAM` | Agent to launch: `claude`, `openclaw` (omit to pick interactively) | — | | `--model` | Model ID to load (omit to pick interactively) | — | | `--port` | Port for the model server | `6767` | | `--api-key` | API key (exported as `OPENAI_API_KEY` and `ANTHROPIC_AUTH_TOKEN`) | `jan` | | `-v, --verbose` | Print full server logs | — | As with `jan serve`, the `--ctx-size`, `--n-gpu-layers`, and `--fit` flags are accepted but currently ignored in router mode. Adjust per-model inference parameters from the desktop app. When no agent or model is specified, interactive selectors are shown. If no models are installed, Jan will automatically download its default foundation model before launching the agent: ``` $ jan launch ━━━ Select Agent ━━━ Choose an agent to launch: > Claude Code — Anthropic's AI coding agent OpenClaw — Open-source autonomous AI agent ``` **Examples:** ```bash jan launch claude # pick a model, then run Claude Code jan launch claude --model qwen3.5-35b-a3b # use a specific model with Claude Code jan launch openclaw --model qwen3.5-35b-a3b # wire OpenClaw to a local model ``` --- ### `jan models` [#jan-models] List and manage models installed in the Jan data folder. ```bash jan models list # list all installed models jan models load # serve a model (alias for jan serve) jan models load-mlx # load an MLX model (macOS / Apple Silicon only) ``` --- ### `jan threads` [#jan-threads] List and inspect conversation threads saved by the Jan desktop app. ```bash jan threads list # list all threads jan threads get # get a thread's metadata jan threads messages # list all messages in a thread jan threads delete # permanently delete a thread ``` --- ## Common Workflows [#common-workflows] **Serve a model for use with any OpenAI-compatible client:** ```bash jan serve jan-code-4b --fit ``` **Launch Claude Code against a local model:** ```bash jan launch claude --model jan-code-4b ``` **Run a model in the background:** ```bash jan serve jan-code-4b --detach ``` **List all installed models:** ```bash jan models list ``` --- ## Troubleshooting [#troubleshooting] For common issues — including Windows-specific problems like `jan` opening the desktop app instead of the CLI — see the [Troubleshooting guide](/desktop/troubleshooting#jan-cli).