Local vs sealed

Two ways to run, one agent#

The CLI is always local software. Only the *model* is optionally remote.

Fully local — nothing leaves your box#

Download an abliterated model, serve it on any OpenAI-compatible local runtime, and point the CLI at it:

# serve your model on a local OpenAI-compatible endpoint, then:
darkroom --model darkroom-local/<your-heretic-model> \
  "find every old wallet I've touched and draft a script to sweep my own dust"

The model, the browsing, and the code all run on your machine. No API, no account, fully offline. This is the most private setup there is.

Sealed API — our compute#

Don't want to host a model? Point it at the Darkroom endpoint:

export DARKROOM_API_KEY=dk-...     # from connecting a wallet at darkroom
darkroom "research X and write me a brief"

This defaults to darkroom/noir running inside attested confidential hardware. You pay per token in crypto; the prompt is sealed to the enclave exactly like the web room.

Presets#

darkroom setup writes two providers into your local model config:

Edit darkroom-local to match the model you pulled, or set DARKROOM_API_KEY to use the sealed API.