README
¶
cLLMHub CLI
The command-line interface for cLLMHub — turn your local LLM into a production API.
What it does
- Download GGUF models from Hugging Face with interactive quantization selection
- Run a local inference daemon with hardware auto-detection (Apple Silicon, NVIDIA, CPU)
- Publish models to the hub so anyone with an API key can use them
- Manage downloaded models — list, search, and delete
- Connect external backends (Ollama, vLLM, LM Studio, llama.cpp, MLX) to the hub
- Authenticate via OAuth device flow, manage credentials
Quick start
# Install (pick one)
npm install -g cllmhub # npm
brew install cllmhub/tap/cllmhub # Homebrew
curl -fsSL https://raw.githubusercontent.com/cllmhub/cllmhub-cli/main/install.sh | sh # shell script
# Authenticate
cllmhub login
# Download and publish a Hugging Face model
cllmhub models --search mistral
cllmhub download TheBloke/Mistral-7B-v0.1-GGUF
cllmhub publish Mistral-7B-v0.1
# Or publish from an external backend
cllmhub publish -m llama3 -b ollama
Installation
npm / npx
# Install globally
npm install -g cllmhub
# Or run without installing
npx cllmhub --help
The npm package automatically downloads the correct pre-built binary for your platform on install.
Homebrew
brew tap cllmhub/tap
brew install cllmhub
Pre-built binaries
Download from your hub's Settings > Downloads page, or grab from GitHub Releases. Available for:
| Platform | Architecture |
|---|---|
| macOS | Apple Silicon (arm64), Intel (amd64) |
| Linux | x86_64 (amd64), ARM64 |
| Windows | x86_64 (amd64) |
Install script
curl -fsSL https://raw.githubusercontent.com/cllmhub/cllmhub-cli/main/install.sh | sh
Build from source
Requires Go 1.22+.
git clone https://github.com/cllmhub/cllmhub-cli.git
cd cllmhub-cli
make build
# Binary is at bin/cllmhub
Cross-compile for all platforms:
make build-all
Commands
Model management
cllmhub models
List downloaded models, or search Hugging Face for GGUF models. Results only include text-generation models.
cllmhub models # List downloaded models
cllmhub models --search mistral # Search Hugging Face
cllmhub models --search "llama 7b"
Flags:
--search, -s Search Hugging Face for GGUF models
cllmhub download <repo...>
Download GGUF model files from Hugging Face repositories. Lists available GGUF files and lets you pick which quantization to download.
For faster downloads and access to gated models, pass a Hugging Face token with --hf-token (it will be saved for future use). Without a token, downloads may be slower and rate-limited.
cllmhub download TheBloke/Mistral-7B-v0.1-GGUF
cllmhub download --hf-token <token> TheBloke/Mistral-7B-v0.1-GGUF
cllmhub download TheBloke/Mistral-7B-v0.1-GGUF TheBloke/Llama-2-7B-GGUF
Flags:
--hf-token Hugging Face token (saved for future use)
cllmhub delete <model...>
Delete one or more downloaded models. Prevents deletion of currently published models and shows freed disk space.
cllmhub delete mistral-7b
cllmhub delete m1 m2 # Use aliases
Daemon
The daemon runs a local llama-server instance and manages model publishing.
cllmhub start
Start the cLLMHub daemon. Engine settings are auto-detected based on your hardware (Apple Silicon, NVIDIA GPU, or CPU) but can be overridden.
cllmhub start # Auto-detect everything
cllmhub start --ctx-size 8192 --flash-attn --slots 2 # Custom settings
cllmhub start --n-gpu-layers 0 --ctx-size 2048 # CPU-only
Flags:
--ctx-size Context size for inference (0 = auto-detect)
--flash-attn Enable flash attention (auto-enabled on Apple Silicon/NVIDIA)
--slots Number of concurrent inference slots (0 = auto-detect)
--n-gpu-layers Number of layers to offload to GPU (-1 = auto, 0 = CPU only)
--batch-size Batch size for prompt processing (0 = auto-detect)
cllmhub stop
Stop the running cLLMHub daemon.
cllmhub status
Show daemon status, including PID, uptime, and currently published models.
cllmhub logs
Show daemon logs.
Flags:
--follow, -f Follow log output
--lines, -n Number of lines to show (default: 50)
Publishing
cllmhub publish
Publish models to the cLLMHub network. Supports two modes:
Daemon mode — publish downloaded GGUF models via the daemon. Auto-starts the daemon if needed.
cllmhub publish llama3-8b
cllmhub publish llama3-8b mistral-7b # Multiple models
Foreground mode — connect to an external inference backend. Keeps a persistent WebSocket connection; your model is online as long as the CLI is running.
cllmhub publish -m llama3-70b -b ollama
cllmhub publish -m mixtral-8x7b -b vllm
Features:
- Auto-reconnect on WebSocket disconnect (up to 5 retries)
- Model server health monitoring
- Heartbeat to keep your provider registered on the hub
- Rate limiting and concurrency control
- Request audit logging
Flags (foreground mode):
--model, -m Model name to publish
--backend, -b Backend type: ollama | vllm | lmstudio | llamacpp | mlx | custom (default: ollama)
--backend-url Backend endpoint URL (overrides default for the backend type)
--max-concurrent, -c Maximum concurrent requests (default: 1)
--log-file Path to audit log file (JSON lines)
--rate-limit Max requests per minute (0 = unlimited)
cllmhub unpublish <model...>
Stop serving one or more published models. The models remain downloaded locally.
cllmhub unpublish mistral-7b
cllmhub unpublish m1 m2
Account
cllmhub login
Authenticate with cLLMHub using OAuth 2.0 device flow. Opens a browser to complete authorization.
cllmhub whoami
Show the currently logged-in user.
cllmhub logout
Revoke credentials on the server and remove the local credentials file.
cllmhub update
Update the CLI to the latest version. The CLI also checks for updates automatically after each command.
Supported backends
For foreground-mode publishing (cllmhub publish -m <model> -b <backend>):
| Backend | Default endpoint | Notes |
|---|---|---|
ollama |
http://localhost:11434 | Default backend, most common |
vllm |
http://localhost:8000 | High throughput, GPU optimized |
lmstudio |
http://localhost:1234 | Desktop app for running local LLMs |
llamacpp |
http://localhost:8080 | CPU-friendly, quantized models |
mlx |
http://localhost:8080 | Apple Silicon optimized via mlx-lm |
custom |
(user-specified) | Any OpenAI-compatible HTTP server |
License
Apache License 2.0 — see LICENSE.
Directories