Course business + PostgreSQL
The Hub owns registration, login, learners, course collections, outlines, revisions, and account isolation.
- Persistent business records
- Each user sees only their own work
- A real domain model you can replace
Agent Builders · runnable reference
See a real Hub before you build your own.
Hub Starter is an Apache-2.0 course-business reference: accounts, private PostgreSQL data, per-user Agent work, selectable AI inference, and Kungfu work control in one locally bound Docker Compose application.
Alpha boundaryThis is a runnable reference for exploration—not a production SaaS, public-ingress, billing, or high-availability claim.
Hub Starter · Docker Compose
One command. No source checkout.
docker compose -f oci://ghcr.io/kungfu-systems/runtime-images/hub-starter:compose-preview up --waitChoose another Web port without exposing the database:
HUB_PORT=9090 docker compose -f oci://ghcr.io/kungfu-systems/runtime-images/hub-starter:compose-preview up --waitUpdate the image without losing your courses.
Run the same Compose application with --pull always. Keep the same HUB_PORT, HUB_PUBLIC_ORIGIN, and COMPOSE_PROJECT_NAME values you used at first start.
docker compose -f oci://ghcr.io/kungfu-systems/runtime-images/hub-starter:compose-preview up --pull always --wait1 · PullFetch the current Buildchain-qualified Alpha application and Hub image.
2 · RecreateReplace changed containers and wait until their health checks pass.
3 · PreserveKeep PostgreSQL, Kungfu state, model files, and credentials in named volumes.
Keep local data: do not run docker compose down -v. To pin or roll back, use an immutable compose-v<VERSION> coordinate from the release list.
What this means for a Builder
The product, the model, and the work runtime are separate authorities.
You can replace each layer deliberately. The reference makes their boundaries visible instead of hiding everything behind one “AI app” label.
Mock, local Qwen, or your endpoint
Mock is ready immediately. A local model is downloaded only after the user chooses it; external OpenAI-compatible inference can be configured separately.
- No model download in the first image pull
- Model binding remains visible in the UI
- Generated content carries provenance
Durable handoff, evidence, and verdict
Kungfu does not generate the course and does not own the customer. It preserves the work truth around Agent actions so another actor can inspect and continue it.
- Bounded work identity
- Evidence and review state
- Continuation beyond one chat
Where the Agent appears
The Agent turns a business intent into reviewable work.
In the course reference, the Agent is not a magic button. It performs a visible sequence, while the application retains the resulting business record and Kungfu retains the work-control facts.
You describe the course
Learner, outcome, constraints, and source expertise enter as business input.
The bound model proposes
AI-produced promises, modules, exercises, and unresolved questions are visibly labelled.
Agents hand off work
Draft, critique, revision, and approval appear as staged actors and observable actions.
Your account keeps the result
The course joins your collection; PostgreSQL owns the record and Kungfu exposes its work truth.
Claim boundary
A starting point, not a finished commercial platform.
Included and runnable
- Course-domain PostgreSQL schema and persistence
- Registration, login, sessions, and account isolation
- Per-user course collections and Agent-produced revisions
- Mock inference plus opt-in local or compatible inference
- Kungfu Agent brief, verification, evidence, review, and seal paths
- Multi-architecture amd64 and arm64 delivery
Not claimed by this alpha
- Public Internet exposure or hardened ingress
- Billing, subscription, email, or organization administration
- High availability, automated backups, disaster recovery, or operational SLOs
- A model that downloads without explicit user choice
- A replacement for your product policy or customer relationship
- Finished integration with every future Kungfu Hub capability