One guarded door to rented compute
Manifold
A local cockpit for launching cloud GPUs, running durable jobs, serving open-weight models, and shutting everything down without losing work or leaking budget.
- Clients
- Next.js dashboard, Tauri desktop app, MCP, and local API
- Control plane
- Python, FastAPI, SQLite, and audited policy guards
- Providers
- Lambda Cloud first, with bounded Google Cloud support
- Runtime
- Supervised containers, SSH tunnels, jobs, and persistent filesystems
The project
What it does
Manifold makes rented GPU infrastructure feel like a local workstation while keeping every expensive or destructive action behind one backend. The UI, desktop shell, MCP tools, and OpenAI-compatible proxy are intentionally thin clients. Budget, ownership, concurrency, data safety, and teardown rules live in the control plane where no client can bypass them.
Work runs as durable job templates rather than fragile shell sessions. A job can rent its own instance, stream logs, synchronize outputs, and terminate the machine when complete. Before any teardown, Manifold attempts to rescue ephemeral files and refuses to proceed when its data-safety policy cannot be satisfied.
Built around
Product principles
Spend that stops itself
Budget, concurrency, idle, active-time, and absolute-lifetime ceilings are enforced by the backend.
Jobs over terminals
Versioned YAML templates supervise serving, fine-tuning, synthesis, media, and custom work.
Agents share the same walls
MCP tools route through the exact policy and audit path used by the human dashboard.
Save before destroy
Termination rescues files first, records the result, and requires an explicit override to burn unsaved work.
System architecture
How the system moves
Every interface converges on the same FastAPI gateway. That gateway owns policy, provider credentials, audit history, job state, and teardown, so adding a new client does not create a new route around the guardrails.
Request to safe compute
-
01 · Interfaces
Dashboard, desktop, MCP
People and agents submit the same typed launch, job, storage, and model requests.
-
02 · Gateway
FastAPI control plane
Authentication, ownership, request validation, and all mutations meet at one door.
-
03 · Guards
Budget, policy, audit
Spend ceilings, concurrency, purpose, attribution, and data safety are evaluated centrally.
-
04 · Providers
Lambda and Google Cloud
Provider adapters launch bounded infrastructure and expose only managed connection paths.
-
05 · Work
Jobs, models, and files
Supervised containers stream logs, survive restarts, and synchronize durable outputs.
-
06 · Teardown
Rescue, verify, stop
Files are saved, resources are terminated, and the final state is written to the audit log.
The invariant is architectural, not conversational: clients request work, but only the guarded backend can spend money or destroy a resource.
Keep exploring