Check it works

Three checks, quickest first. If any of them fails, waired doctor is the one command to run next.

1. Look at the Waired icon

Click the Waired icon — in the menu bar on macOS, next to the clock on Windows and Linux.

screenshot pendingThe Waired menu showing the signed-in account, a Connected status, and an Inference submenu naming the active model.public/img/app-ready.png

Signed in, connected, and a model loaded.

What good looks like: your account address at the top, a connected state, and — on a computer that runs models — Inference naming the active model.

2. Ask it something

waired infer "say hi"

What good looks like: a greeting, within a few seconds.

This goes straight to the model on this computer, so it works even if nothing else on your network is reachable. It is the fastest way to prove the model itself is alive.

To see where the answer came from and why:

waired infer --explain "say hi"

3. Read the full status

waired status --observability

What good looks like, on a computer that runs models:

Observability:
  Device:   my-desktop  uptime: 1m23s
  Engine:   ready (model=qwen3.8-27b, 0/4 slots used, inflight=0)
  Share:    enabled   Paused: no
  Mesh:     2 enrolled / 1 reachable / 1 ready
  Last:     (no inference yet)

Line	What to check
Engine	ready means a model is loaded and can answer. not ready usually means it is still downloading — see No answer comes back.
Share	enabled means your other computers may use this one’s model. disabled has two possible causes: sharing is off on this computer (waired share off, or the app is closed), or the Waired console has taken it out of your mesh. waired share status says which.
Paused	no means Waired is handling requests. yes — run waired resume.
Mesh	enrolled / reachable / ready across your computers. On a laptop, this is where you confirm the computer running your model is up.
Last	Nothing on a fresh install. Once you have asked something it summarises that request: ttft is how long you waited before the first word appeared, latency is how long the whole answer took. Ask a follow-up in the same session and ttft usually drops sharply, because the model is reusing what it already read.

Add -o json for a machine-readable version.

On a laptop that does not run models

Engine: not ready is correct there — that computer is not supposed to host a model. What matters is the Mesh line showing at least 1 ready, which means one of your other computers is available to answer.

Then try it for real

The real test is your own tools:

Claude CodeAsk it something and read the footer — it names the model that answered.

OpenCodeOne command to connect it.

OpenClawOne command to connect it.

A chat appAny OpenAI-compatible client.

Something looks wrongFind the symptom, get the fix.