From repositories to a local AI community
OpenFace v0.1.0 starts with a deliberately small premise: the durable objects of an AI community should remain ordinary Git repositories and ordinary applications.
Keep the source of truth boring
Forgejo owns files, commits, tags, Issues, Pull Requests, users, organizations, teams, stars, and forks. OpenFace adds a discovery and runtime layer without inventing a second repository database. Topics classify Models, Datasets, Spaces, Characters, Benchmarks, Skills, MCPs, Prompts, and Knowledge.
That boundary makes the catalog portable. A clone is still a clone, a tag is still immutable, and rebuilding the portal does not rewrite project history.
Treat a Space as an application
A root Dockerfile is the common contract. Framework-specific samples—Gradio, static HTML, React, Vue, Next.js, Streamlit, FastAPI, and Node.js—listen on port 7860, while the runner handles build, lifecycle, and proxying.
Public applications can be launched without a Forgejo session. Destructive or configuration-changing controls remain owner-only. Existing services can instead use external_url, and static documentation can use OpenFace Pages.
Separate the control plane from compute
The core stack can live in an LXC while GPU workers claim jobs over the LAN. The worker announces capabilities and free VRAM; the scheduler chooses a compatible node without exposing worker credentials to the browser. This keeps an inexpensive CPU host useful while leaving room for RTX-class acceleration.
Make automation leave evidence
The maintenance flow is not a hidden cron job. A coordinator classifies work, delegates to a specialist identity, runs Claude Code /goal, opens a Pull Request, and requests an independent review against a fixed commit SHA. Auto-merge is fail-closed when evidence, labels, or review state is missing.
Verify the surface people actually use
OpenFace’s visual suite captures desktop and mobile routes across Standard, Solarpunk, and Cyberpunk themes. It scrolls long pages, exercises menus and controls, validates Markdown and Mermaid rendering, and stores the screenshots beside the implementation evidence.
The result is not just a catalog and not just a runner. It is a local collaboration surface whose data, applications, automation, and verification can all be inspected.
Continue with the full v0.1.0 release notes or build your first deployment.