How it works
Darkroom is an uncensored model running inside sealed hardware, paid for in crypto. Three properties, two surfaces.
The one-paragraph version#
Darkroom gives you an uncensored AI brain that runs inside sealed hardware and that you reach without an account. Those are three independent properties, and each one is enforced by something you can check — not a promise in a privacy policy.
Three properties#
Uncensored#
The models are abliterated open-weight models: the refusal reflex has been edited out of the weights, not prompted around. They answer the question instead of lecturing you about it. See Abliterated models.
Private#
Your prompt is only ever readable inside a sealed room — confidential hardware that decrypts the request, runs it, and returns the answer without writing your text anywhere persistent. No logs, no telemetry, no human in the loop. See Sealed rooms.
Unlinkable#
There is no email and no password. You authenticate with a wallet signature and pay in crypto. Nothing ties a session back to your name. See The privacy model.
Two surfaces, same brain#
You can use Darkroom two ways:
| Surface | Runs where | Best for |
|---|---|---|
| The room (web) | our sealed hardware, in your browser | zero setup, chat, vision, quick answers |
| The agent (CLI) | your machine | operating on your own files, shell and the open web |
The web room is the fastest way in — open it, connect a wallet, talk. The CLI is a local agent that *does* things on your own hardware, with the model running either fully offline or against the same sealed endpoint the room uses. Start with the room quickstart or install the CLI.
What you can verify#
You don't have to trust any of the above. The room runs in attested hardware with a fixed, public fingerprint measured at boot. Anyone can check that the fingerprint matches and that the prompt path wasn't tampered with. See Verify it yourself.