@offgrid/gateway

The local, OpenAI-compatible gateway behind Off Grid AI — run open models (text, vision, image, voice, speech) entirely on-device, behind one API. No cloud, no account, no API key.

It runs two ways from one codebase:

• Embedded — Off Grid AI Desktop injects its Electron paths and runs the gateway in-process.
• Standalone — offgrid-gateway (or the Docker image) runs it headless on its own box.

# standalone (Node)
OFFGRID_DATA_DIR=~/.offgrid OFFGRID_BIN_DIR=/opt/offgrid/bin npx @offgrid/gateway
# → http://127.0.0.1:7878  (OpenAI-compatible: /v1/chat/completions, /v1/images, …)

# Docker (mount your models + platform binaries)
docker run -p 7878:7878 \
  -v /models:/data/models -v /opt/offgrid/bin:/bin/offgrid \
  -e OFFGRID_DATA_DIR=/data -e OFFGRID_BIN_DIR=/bin/offgrid \
  ghcr.io/off-grid-ai/gateway

Configuration

All host-specific paths resolve through one seam (runtime-env):

Env var	Meaning
OFFGRID_DATA_DIR	writable dir for models, caches, generated output
OFFGRID_BIN_DIR	dir holding the platform binaries (llama-server, sd-cli, whisper, ffmpeg)
OFFGRID_RESOURCE_DIR	dir holding bundled resources (tts worker, …)
OFFGRID_GATEWAY_PORT	listen port (default 7878)

Embedded hosts call configureRuntime({ dataDir, binRoots, resourceDirs }) instead.

Status

v0.1 — the host-agnostic config seam + deployment surface (CLI, Docker, CI for GitHub Packages + GHCR) are in place. The OpenAI-compatible inference handlers (chat/vision/image/audio/embeddings) are migrating from the desktop runtime; until then the standalone server exposes /healthz and /v1/models.

License

AGPL-3.0-only. © Off Grid AI / Wednesday Solutions, Inc.