video-factory

GPU infrastructure control for Pi. Install one package and your agent can provision, deploy, run, monitor, and destroy GPU workloads on whatever cloud infrastructure you have access to — with video generation (MiniMax H3, Wan 2.2, HunyuanVideo, Seedance) as the first-class workload.

pi install npm:@lvrged/video-factory

Then, inside Pi:

I need HunyuanVideo running on a GPU. Find the cheapest suitable provider, deploy it, expose an endpoint, test it, and make it available to me as a tool.

The agent figures out the rest — from docs/pricing/ for provider comparison, from the skills for the how-to, and from the ledger for cost control.

What you get

• gpu_* tools — the normalized agent interface: gpu_setup, gpu_status, gpu_ensure, gpu_provision, gpu_register, gpu_run, gpu_job_finish, gpu_jobs, gpu_spend, gpu_destroy, gpu_model_add, gpu_workflow_add, gpu_policy.
• /gpu commands — the human dashboard: setup, status, deploy, jobs, spend, stop, policy.
• Skills — the knowledge: gpu-setup, gpu-ops, comfyui, model-deployment, video-models, provider-adapters.
• Docs — pricing snapshots for RunPod, Vast.ai, Modal, Lambda, Prime (docs/pricing/) and per-provider playbooks (docs/providers/).
• A persistent registry + job ledger in <project>/.pi/gpu/ — the agent restarts tomorrow and still knows H3 is deployed, what it costs, and which workflow version made which video.

Quickstart

1. First run = onboarding. The extension greets you with /gpu onboard — the H3 production economics (Turbo 4-8 step workflow, provider lanes, cost per finished minute) and the first-deployment flow. Or just say: "set me up to make H3 videos cheaply" — the agent runs the gpu-onboarding skill.
2. /gpu setup (or ask the agent to run gpu_setup) — detects provider CLIs, installs missing ones with your permission, checks auth.
3. Ask for a model: "deploy h3 on the cheapest provider" — the agent runs gpu_ensure, searches live prices, provisions, installs ComfyUI + weights (with the optimized Turbo/Sage/INT8 workflow), health-checks with a test generation, and registers the deployment.
4. Generate: "run h3-image-to-video on slide-042.png with 'slow push-in', 5 seconds" — gpu_run opens a ledger entry; artifacts sync off the pod before shutdown.
5. "How much did the last 100 videos cost?" — gpu_spend answers from the ledger. "Regenerate everything from workflow v3" — gpu_jobs finds them.

Provider coverage

| Provider | Interface | Kind | Notes | | --- | --- | --- | --- | | RunPod | runpodctl + API | managed instances | community (cheap) / secure pools, templates | | Vast.ai | vastai | marketplace | cheapest lane; search live before every deploy | | SaladCloud | REST API | containers on distributed GPUs | batch pricing — the economic outlier; bring your own image | | Modal | modal | serverless | zero idle cost, cold starts | | Lambda | REST API | premium on-demand | no egress fees; clusters need commitment | | Prime | prime | decentralized marketplace | dynamic pricing, spot, parallel bids |

New providers don't need a new release: if its CLI is agent-friendly, the provider-adapters skill teaches the agent how to operate it.

Using the same skills with Claude Code (no Pi)

The repo ships a parallel skill set in claude-skills/ for people who don't run Pi — same knowledge, adapted to bash-only operation (no gpu_* tools), sharing the same .pi/gpu JSON registry so Pi and Claude users can operate identical infrastructure:

cp -r claude-skills/* /path/to/your/project/.claude/skills/

Claude Code picks them up automatically (gpu-setup, gpu-onboarding, gpu-ops, comfyui, model-deployment, video-models, provider-adapters). The onboarding skill walks a new Claude user through the same H3 economics card.

Architecture

See docs/ARCHITECTURE.md (the core abstraction is the deployment, not the provider) and docs/REGISTRY.md (state file schemas).

                Pi Agent
                    │
          video-factory extension
                    │
           provider capability tables
        ┌──────────┼──────────┬──────────┐
        │          │          │          │
     RunPod     Vast.ai    Modal     Lambda/Prime
        │          │          │          │
     CLI/API    CLI/API     CLI        API

The spend policy (read this)

Extensions run arbitrary code — this one executes provider CLIs and spends real money. The trust boundary is .pi/gpu/policy.json:

{ "ceiling_per_job_usd": 5, "ceiling_daily_usd": 40, "ceiling_monthly_usd": 400, "confirm_above_usd": 1, "idle_shutdown_after_min": 30 }

Provisioning above confirm_above_usd always asks you in the UI, and ceilings cap what the agent may do without checking in. Adjust with gpu_policy or /gpu policy. Review this package's source before installing — it's yours.

Development

npm install          # dev deps (typescript, pi types)
npx tsc --noEmit     # typecheck
pi -e .              # load the extension from this dir in a scratch session

License

MIT