06AI infrastructure · 2026 Interactive

Flock

A GPU cloud that lives in one PC.

Flock runs Modal-style apps on a single RTX 5090 at home: the same decorators, the same dashboard sections, no per-second bill. The console below simulates its scheduler. Send a render and watch the warm container, the GPU and the electricity meter respond.

Role: Platform design, API server, Python SDK and CLI, dashboard, container images
Platforms
Stack: Python · FastAPI · Docker · NVIDIA Container Toolkit · Tailscale
Status: In use on a home RTX 5090; render farm in Phase 1

How it was made

Flock console Warm T+00:00.0 Simulated timings

The Flock console needs JavaScript. It simulates calls, a warm container and GPU telemetry.

Run .remote() once, then .map() a batch.Stop the container to see a cold start.Timings are compressed: the real scaledown window is five minutes.

1RTX 5090 serving every app, call, endpoint and notebook
12Modal features mirrored, from warm classes to queues and dicts
0Ports open to the internet: access is over Tailscale
£/kWhCost per app measured from the GPU's real power draw

The briefModal's ergonomics, without the meter

Serverless GPU platforms are a joy to use and expensive to leave running. Flock keeps the developer experience, with @app.function, @app.cls, @flock.enter, .remote(), .spawn() and .map(), and moves the hardware home. An existing Modal ComfyUI app ports by changing the import.

ArchitectureEverything is a container

The platform itself is a container, and so is every function call, warm class, endpoint, notebook and sandbox. The server drives Docker through its socket with --gpus all, builds images once from a Modal-style image builder and caches them, and keeps model weights in named volumes.

README.md · what runs on the GPU boxdiagram

your laptop                            GPU box
───────────                            ──────────────────────────────────────
flock.py (SDK + CLI) ──HTTPS/Tailscale──▶ flock-server  :8000  (API + dashboard)
browser ──────────────────────────────▶      │ docker.sock
                                             ├─ flock-fn-<call>    one-shot calls
                                             ├─ flock-cls-<app>-…  warm @app.cls containers
                                             ├─ flock-ep-<name>    endpoints  :9000+
                                             ├─ flock-nb-<name>    JupyterLab :9000+
                                             └─ flock-sb-<name>    sandboxes

ProductA dashboard you already know

Apps, functions, containers, endpoints, secrets, storage, notebooks and workspace metrics, laid out like Modal's so there is nothing to learn. Deploys are versioned with rollback. Endpoints serve ComfyUI, vLLM, Ollama or any container in one click. Electricity is measured from the GPU's power draw and shared between the apps running at each moment.

NextFrom one GPU to a farm

Phase 1 of a community render farm is built on the same pieces. Providers lend idle GPUs with one command and a browser approval. Render workers run with no network at all, every connection is outbound from the machine, and a lease system requeues only unfinished frames when a node drops, so a gamer reclaiming their GPU never loses anyone's work.

What this means for your project

GPU infrastructureDocker, NVIDIA drivers, CUDA and Blackwell-ready images for AI workloads.
Platform engineeringAn API server, SDK, CLI and dashboard that feel like a product.
AI deploymentComfyUI, vLLM and Ollama endpoints, warm containers and queues.
Secure by defaultPrivate networking over Tailscale, token auth and isolated workers.