Atlas

Atlas turns your machine into a private AI host. It powers Curie's local path and answers any client already fluent in /v1/chat/completions. Private by default, free, forever.

Start Atlas, point any OpenAI-compatible client at localhost:8787, that is the whole integration.

# Install Atlas via Toolbox
toolbox install atlas

# Pull a model and serve
atlas pull pulsar
atlas serve

# Verify
curl http://localhost:8787/v1/models

curl http://localhost:8787/v1/chat/completions \
  -H 'Authorization: Bearer atlas-local' \
  -H 'Content-Type: application/json' \
  -d '{
        "model": "pulsar",
        "messages": [
          {"role": "user", "content": "Hello."}
        ],
        "stream": false
      }'

Atlas exposes a strict subset of the OpenAI HTTP surface. Streaming follows the same SSE protocol. Endpoints not yet implemented return 501 with a typed body, never a silent fallback.

• POST /v1/chat/completions, chat with streaming.
• POST /v1/embeddings, dense embeddings.
• GET /v1/models, list available local models.
• GET /healthz, liveness, returns the loaded model digest.

Atlas ships with Pulsar, the Scalaris vision / agent / text model, and lets you install your own models through the same manager. You ask for a model by name; Atlas loads it and serves it locally, nothing leaves the machine.

All configuration lives in a single TOML file under your platform's config directory. The CLI prints the path with `atlas config path`. Settings are hot-reloaded on file change.

{
  "port": 8787,
  "defaultModel": "pulsar",
  "egress": "deny"
}

If `atlas serve` exits with a non-zero status, check the model first. Most failures are upstream model files, not Atlas itself.

• Port 8787 already in use, pass `--bind 127.0.0.1:8788` or kill the previous instance.
• Out of GPU memory, Atlas falls back to CPU only when `--allow-cpu` is set.
• Model not found, run `atlas pull <model>` to fetch and verify the digest.