See what your machine is actually doing with Ollama — not what a browser can guess.

Install & run

On npm since August 2026 — one command, nothing to clone, zero dependencies beyond Node itself:

npx @opensourcesai/cmdcenter serve      # dashboard at http://127.0.0.1:7717
npx @opensourcesai/cmdcenter            # one-shot human-readable report
npx @opensourcesai/cmdcenter --json     # the full report as JSON

Every release is published from public CI with npm provenance — the registry can prove each tarball was built from a public commit of the repository, and there is no other publish path. Prefer reading before running? Clone the repository and run the same commands as node src/cli.js … — npx runs exactly that source.

You need

Stopping it, and removing it. Press Ctrl+C in the terminal running the server. It installs no service and registers no startup entry. Since the chat surface landed (August 2026) it keeps exactly one directory of its own data — your conversations and this machine’s measurement history — at %LOCALAPPDATA%\osai-cmdcenter on Windows, ~/Library/Application Support/osai-cmdcenter on macOS, or ~/.local/share/osai-cmdcenter on Linux. Delete that directory and the tool is gone without a trace beyond npx’s ordinary package cache; individual conversations can also be deleted from the Chat view itself, behind a confirm. Nothing in that directory is ever transmitted or read into a diagnostic capture.

What it looks like running

What it reports

What this adds to ollama ps

If you already use ollama ps, keep using it. It ships with Ollama, it costs nothing to run, and it answers its question honestly: what is loaded right now, and how much of it landed on the GPU. Nothing here replaces it.

What it can’t tell you is whether that split was avoidable. Command Center reads the same runtime numbers and sets them beside the three facts that explain them.

How much VRAM you actually have. Read from every source your platform exposes rather than one. Where two sources disagree — and on Windows they routinely do, because one common interface tops out at about 4 GB and will happily report that for a 12 GB card — both figures are shown and the disagreement is named, instead of being quietly resolved to whichever number looked cleaner.

Whether your models fit it. A committed catalog snapshot is graded against your measured memory — comfortable, tight, runs-but-offloaded, or too large — with the quantization that fits and the exact ollama run command for it. A command appears only where the catalog carries a real registry-verified tag; the tool will not synthesize one, because a copy-pasteable command that 404s is worse than no command.

Residency, in context. A model sitting at 60% in VRAM is a fact ollama ps reports too. The useful version is that number next to the card it’s running on, the headroom left, and the quantization that would have fit whole.

Running it on a headless machine

Plenty of Ollama setups have no desktop at all — a spare tower in a cupboard, a home server, a rented Linux box — and you drive them from a laptop. That works, and the way it works is deliberately inflexible in one specific respect.

There is no flag to put the dashboard on your network. The bind address is the loopback literal 127.0.0.1, written into the source with a test that fails the build if it ever changes. --port moves the port; nothing moves the address. So the usual way a local dashboard ends up quietly reachable from the rest of the house — a well-meaning --host 0.0.0.0 — is not available to you, and is not available to anyone who talks you into it either.

The supported way in is an SSH tunnel, which keeps every guarantee above intact. Start the server on the machine Ollama runs on:

nohup npx @opensourcesai/cmdcenter serve --port 7717 < /dev/null > /tmp/cmdcenter.log 2>&1 & disown

The < /dev/null is not decoration, and it is the part people leave out. Without it the server keeps the SSH channel open and your terminal appears to hang — even with nohup and disown — which reads as a failure when the server has in fact started perfectly. Redirecting standard input closes the channel and gives you your prompt back. The log file is where any startup error will be.

Then, from your laptop, in a new terminal — not inside the SSH session you just used:

ssh -N -L 7717:127.0.0.1:7717 your-server

Leave that running and open http://127.0.0.1:7717/ in your browser. There is no token to copy or paste: the server hands it to the page itself. The dashboard is still bound to loopback on the server, still requires the per-session token on every request, and still checks the Host and Origin headers. The tunnel changes where the loopback connection comes from, not what is guarding it.

Use the same port number on both sides. The Origin check is pinned to the port the server was started on, so forwarding local 7718 to remote 7717 leaves you with a dashboard that loads and reads fine but refuses its two actions. Matching the numbers avoids the whole question.

When it doesn’t work

Four things that have actually gone wrong, and what each one really means. None of them are what the error message suggests.

channel 3: open failed: connect failed: Connection refused, repeating. The tunnel is fine. That message is the tunnel reporting that it arrived and found nothing — nothing is listening on 127.0.0.1:7717 at the far end, which nearly always means the server isn’t running. Check the log you redirected it to.

Permission denied when the tunnel tries to bind the local port on Windows. Not a permissions problem, whatever the wording says: something on your own machine already holds 127.0.0.1:7717. An earlier tunnel you never closed is the usual culprit. Close it rather than forwarding to a different local port — see the matching-ports note above.

A host-key prompt for a machine you are already logged into. The tunnel command was run inside the SSH session instead of on your laptop, so the server is being asked to SSH to itself. Open a second terminal locally and run it there.

“Ollama not detected” while Ollama is plainly running. The Ollama Windows app sets OLLAMA_HOST=0.0.0.0, and that one variable means two incompatible things: to the Ollama server it is the address to listen on, to a client it is the address to connect to. 0.0.0.0 is a sensible instruction to listen everywhere and a meaningless place to dial, so a client that takes it at face value fails instantly on a machine where Ollama is working perfectly. Command Center already handles this — a wildcard address falls back to loopback, keeping any non-default port you set. It’s written up here because the same variable breaks other local AI tools the same way, and the symptom never points at the cause.

What this isn’t — on purpose

A measured chat surface, not a chat product. This page used to say “not a chat client”, and until August 2026 that was true. It now has a chat view — but the point is the instrument: every reply arrives with its own measured figures and a verdict on whether the machine kept the fit engine’s promise. If you want a full-featured chat frontend, keep the one you have; this one exists to tell you what your machine actually did while it answered.

Not a fleet or enterprise tool. It reports on the one machine it runs on, and it never compares your machine to anyone else’s.

Still no installer. It ships as plain source on npm — npx runs it directly on Node. No OS installer, no code signing, no auto-update; releases are published only from public CI, with provenance the registry can verify.

Exactly two actions, and inference — nothing else. Load and unload remain the entire action surface. The chat view additionally relays inference to the AI runtimes on your machine — Ollama, and optionally a local llama.cpp server — and that is the entire list: it talks only to AI runtimes on this machine, never to the internet, and nothing else in Ollama’s API is reachable through it.

Found a bug, or stuck?

Detection is the genuinely hard part of a tool like this, and machines differ in ways no test suite anticipates — every platform quirk it handles today got found on somebody’s real hardware. A number that looks wrong is worth reporting, not working around.

Next steps