cli.md

title	CLI reference
description	Manage an inference server on one GPU machine or several

Usage

llm-helper

Description

llm-helper (formerly vllm-helper) -- serve large models with vLLM, mlx-lm or llama.cpp
on one GPU machine or several (models, several at once if memory allows,
sharding, replica pools, recipes, optional Coder Agents sync).

Subcommands

Name	Purpose
adopt	Detect an already-running vLLM server and save it as a managed profile
coder-sync	Manually sync a profile's Coder model entry
default	Show or set the model Coder uses by default
doctor	Diagnose hardware, running servers, integrations, and config drift
down	Stop serving a profile (the only one running when none is named)
edit	Open a profile in $EDITOR and validate the result
engines	List the inference engines a profile or recipe can run on
hardware	List hardware definitions and report this machine's hardware
inspect	Show a model's license, weight size against this machine, and detected flags
install-gguf	Install the plugin that lets vLLM load GGUF models
list	List saved model profiles
logs	Show a running model's server log, following by default on a terminal
metal-limit	Show or raise how much unified memory macOS lets the GPU use
migrate	Move a model started by an older release to the per-model services
model-types	List what specific model architectures need to run correctly
new	Pull a new HuggingFace model, save it as a profile, and start it
node	Register machines by HTTP URL and SSH target
pool	Run a router in front of replicas serving the same model
recipes	Tested launch configurations for known-good models
remove	Delete a profile, and its Coder model entry when sync is configured
restart	Re-apply a running profile and restart its server
router	Serve every running model behind one port
search	Search HuggingFace and pick a model to pull and run
serve	Run a web panel for this machine
set	Change profile settings from the command line
setup	First-time setup: detect hardware and the inference server, write config, offer to install the service
shard	Split one model across machines (Ray or vLLM multiprocessing)
spark	DGX Spark helpers
status	Running models, and whether each endpoint responds
up	Start serving a profile alongside anything already running
upgrade	Update llm-helper itself from a published release
version	Print the version