ow

ow puts the models on your computer into the OpenWeights pool. It finds the local model server you already run, registers its models, then stays open and takes jobs.

It does not install Python, a model, or a GPU stack. It is one file, with no dependencies.

The same steps, with copy buttons, are on owterminal.com/inference?door=host.

Install

macOS, Linux, or Windows with WSL. Node 22 or newer. A model server already running on this computer.

curl -fsSL https://raw.githubusercontent.com/dvidia-inference/ow/main/install.sh | sh
export PATH="$HOME/.ow:$PATH"
ow models

The same installer is on the site: curl -fsSL https://owterminal.com/ow-install.sh | sh

Both download the current ow.mjs from owterminal.com, check it parses, and save it to ~/.ow.

Your model server

ow talks to any local server with an OpenAI-style /v1 API. Point LITELLM_BASE at it. The name is historical; it is not only for LiteLLM.

Server	LITELLM_BASE
LiteLLM	http://127.0.0.1:4000 (the default)
Ollama	http://127.0.0.1:11434
LM Studio	http://127.0.0.1:1234
llama.cpp llama-server	http://127.0.0.1:8080
vLLM	http://127.0.0.1:8000

A trailing /v1 is fine. If the server wants a key, set LITELLM_KEY. It is sent only to your server.

What the first run does

ow

1. It asks your server which models it has. It waits two seconds.
2. It registers each new model and prints its public node code.
3. It stays open and takes jobs. Leave the window open.

Run it again and it prints the same codes. It does not register a model twice.

While it is open, ow:

• reads the model list every minute, and registers new ones
• times each model with a short "hi" every five minutes
• runs a pool test, and renews it every 12 hours
• asks for work every two seconds

The bottom line says waiting, working, done, failed, or busy GPU is full.

In another terminal, ow status says READY once the pool test passes. ow link prints a one-time code. Enter it at owterminal.com/account/machines within 10 minutes to link this machine to your account.

Stop with Ctrl+C. The pool marks the machine offline after three minutes without contact. ow serve resumes with the same keys.

Commands

ow                 look at this computer, join, then stay open
ow models          list the local server's models, without joining
ow join            register the models, then stop
ow serve           stay open and take jobs, if already registered
ow status          read pool readiness
ow link            get a one-time account pairing code
ow check           request a pool test
ow check status    read the pool test result

The site's flow is ow models, then ow join && ow serve.

Knobs

Set these before ow, in the same terminal.

	Default
LITELLM_BASE	http://127.0.0.1:4000	Your model server.
LITELLM_KEY	empty	Only if your server asks for a key.
OW_MODELS	every model	Exact model IDs, comma separated. Up to 24.
OW_PRICE	suggested	Cents per million tokens, 1 to 5000. 150 is $1.50.
OW_PRICES	empty	Per-model prices as JSON: '{"qwen2.5:7b":100}'.
OW_SLOTS	1	Jobs at once, 1 to 8. For servers that batch, like vLLM or llama-server.
POOL_URL	https://owterminal.com	The desk.
OW_HOME	~/.ow	Where the keys are saved.

OW_MODELS=qwen2.5:7b OW_PRICE=150 ow

Without OW_PRICE, the desk suggests a price for each model and ow join prints it. Prices apply when a model is first registered. A model already registered keeps its price.

Leave OW_WAIT and OW_RELAY unset. Held claims are off by default, and the relay is present but not enabled yet.

Keys

~/.ow holds two key files. Both are private to your user (600). Do not paste them anywhere.

• nodes.json: one secret per model. The secret is how the desk knows this machine.
• enckey.json: an X25519 key pair for sealed prompts.

Some prompts arrive sealed. On first join, ow makes the key pair and sends the desk only the public half. A sealed prompt is decrypted with the private key, in memory, only while its job runs. It is not written to disk. This machine still sees the prompt, because it has to run it. If a prompt does not decrypt, the job is reported failed.

Back up ~/.ow. If you lose nodes.json, the next run registers the models again with new codes.

What it sends

To the desk: each model's ID, the GPU name and memory (from nvidia-smi, or the Apple chip name on a Mac), a measured speed, whether it is busy, the slot count, the public key, and each job's output and token counts.

To your server: the jobs, as chat completions.

The node secrets go only to the desk. LITELLM_KEY goes only to your server. The wire format is in PROTOCOL.md.

When it stops

What you see	What to do
LiteLLM is not running on this computer.	Start your model server. If it is not on port 4000, set LITELLM_BASE.
LiteLLM is running, but the password was refused.	Set LITELLM_KEY.
It has no models loaded.	Load a model, then run ow again.
fetch failed	Nothing answered. Check that your model server is up at LITELLM_BASE and that this computer can reach owterminal.com.
A selected OW_MODELS ID is missing from this server.	Run ow models without OW_MODELS and copy the exact IDs.
Choose up to 24 models with OW_MODELS.	List the ones to share in OW_MODELS.
OW_PRICE must be integer cents from 1 to 5000.	Use whole cents. 150, not 1.50.
OW_PRICES rates must be integer cents from 1 to 5000.	Same rule, inside the JSON.
Already open in another window. / This worker is already running.	Use that window. Do not start a second one.
Nothing joined. Run: ow join	Run ow join first.
Could not read saved host keys.	Restore ~/.ow/nodes.json from a backup. Do not delete it.
Could not read the saved encryption key.	Restore ~/.ow/enckey.json from a backup.
busy GPU is full	nvidia-smi shows more than 92% of GPU memory in use. ow takes no jobs until it drops. Servers that reserve memory up front can trip this.
ow status stays on CHECKING	Keep ow serve open. Read the result with ow check status. Tests run at most once every 10 minutes.
Node.js 22 or newer is required.	Install Node 22 from nodejs.org, then run the installer again.

Where it sits

this computer
  model server     LiteLLM, Ollama, LM Studio, llama.cpp, vLLM
  ow               looks, joins, stays open
        |
        v
owterminal.com     the desk. the code, the price, the jobs

Do not point your model server back at owterminal.com. ow calls your server. Your server must not call the pool.

Two computers, each with its own server, each run their own ow.

Updating

Run the installer again, then restart ow serve. It replaces ~/.ow/ow.mjs and keeps your keys. Changes are in CHANGELOG.md.

License

MIT