feat(mcp): MPP-gated deploy flow — agents pay to host persistent services

[hellno]

Summary

Part 2 of MPP compute on Jack Cloud. Part 1 (execute_code, PR #27) ships instant sandboxed code execution. This issue covers persistent hosting: an agent pays via MPP and gets a live URL that stays up.

Context

• PR feat(mcp): MPP payment sessions + execute_code tool #27 adds execute_code MCP tool with MPP payments (live, verified)
• gakonst tweet: "desperately need MPP-enabled compute & hosting"
• /execute handles the "compute" half. This issue handles the "hosting" half.
• Design docs: ~/.gstack/projects/getjack-org-jack/hellno-machine-payments-design-*.md

What this enables

Agent calls deploy_project MCP tool without a human account. Jack returns an MPP session challenge. Agent opens a Tempo payment channel ($5 min deposit). Agent deploys code, gets a live URL. Subsequent deploys (redeploys, iterations) deduct from session balance. Services suspend when balance hits zero.

Architecture

Agent → MCP Worker → deploy_project tool
                      ├── chargeSession() → ComputeSession DO
                      │   ├── No session → 402 MPP challenge (session intent)
                      │   └── Valid voucher → deduct $0.05 per deploy
                      ├── bundleCode(files) → esbuild-wasm
                      ├── Create project via control plane API (ephemeral org)
                      ├── Upload deployment via FormData
                      └── Return { url, deployment_id, expires_at }

Key decisions needed

1. Bundler location: Extract apps/mcp-worker/src/bundler.ts to packages/bundler/ so template and MCP worker share one implementation? Or duplicate?
2. Ephemeral orgs: MPP sessions need a control plane org to own projects. Create ephemeral orgs with no human owner? Or reuse existing org model with a machine flag?
3. Deploy TTL: 24h auto-expire? Or keep alive until session balance exhausted? (Design doc says "until balance exhausted with 7-day inactivity timeout")
4. Runtime metering: Free deploys, metered runtime (requests + CPU). Requires dispatch worker changes. How much latency does DO balance checking add? (Design doc targets <5ms p50)
5. Cleanup: Cron to garbage-collect expired/abandoned deployments from dispatch namespace

Implementation pieces

• deploy_project MCP tool with MPP session gating
• Ephemeral org creation in control plane for MPP sessions
• Session-to-project cardinality tracking in ComputeSession DO
• Extract bundler to shared package (DRY with MCP worker)
• Dispatch worker metering integration (Phase 2)
• Auto-suspend on zero balance (Phase 2)
• Cleanup cron for expired deployments (Phase 2)

Pricing model (from design doc)

Resource	Rate
Deploy / redeploy	$0.05 per deploy
Worker requests	$0.50 / million
CPU time	$12.50 / million ms
D1 reads/writes	$0.25-1.00 / million

Deploys have a small fixed cost. Runtime is metered. Session minimum $5 USDC.

Dependencies

• PR feat(mcp): MPP payment sessions + execute_code tool #27 merged (execute_code + ComputeSession DO + MPP auth)
• Control plane support for ephemeral orgs
• Bundler extraction (or accept duplication for V1)

References

• gakonst tweet
• MPP spec
• Design doc #1 — compute sessions
• Design doc #2 — mpp-compute template (superseded by platform approach)