Ideas queue up behind one card
Video generation is measured in minutes per shot. Testing ten variants of a scene means an evening — so the tenth variant is never tested, and you ship the third one.
Every idle gaming PC in the studio is a render node. Hive turns them into one queue that a person — or an AI agent — can load with work, and keeps every machine invisible to the outside world.
A studio that generates video hits the same wall every time: the machine renders one job while eleven ideas wait. Cloud GPUs solve it with money; Hive solves it with the hardware you already bought.
Video generation is measured in minutes per shot. Testing ten variants of a scene means an evening — so the tenth variant is never tested, and you ship the third one.
ComfyUI has no authentication of its own. Instances left open on the internet are found and taken over within days — there was a botnet campaign doing exactly that in April 2026. Sharing a friend's PC by opening a port means putting their home network on the table.
An LLM writes fifty workflow variants in a minute. Without a queue it cannot use them: someone has to sit and feed a GUI. The bottleneck stops being the GPU and becomes the human in the middle.
Hive is deliberately small in concepts. There is a coordinator, there are machines, there are senders, and there is a panel where all of it is visible.
One server holds the queue, the scheduler, authentication and the drone registry. It is a control plane only: multi-gigabyte artifacts never travel through it.
A tray app for Windows and macOS. It installs and owns its own ComfyUI, opens an outgoing connection to Queen, and takes work only when the owner is idle and the GPU is free.
A person through the Hive panel, or an AI agent through MCP: submit, poll, fetch. Claude does not need a browser and does not wait — it submits a hundred jobs and comes back for the results.
The colony, the queue, live progress on every job, the gallery of what came out, and the admin side: join codes, keys, quotas, the fleet's version map.
Anything can dispatch a job. The difference shows up on the second day: when a machine sleeps mid-render, when two drones have different model versions, when a workflow fails on exactly one card.
A job is leased to a drone and held by progress heartbeats, not by a fixed timeout. A long render is not stolen from a working machine; a silent machine gives the job back in seconds.
Capability is decided by SHA-256 of every model and pinned commits of custom nodes. Two files called the same thing are not the same thing — and a silently wrong result is worse than a failed job.
A job that fails the same way on two different drones goes to dead-letter instead of touring the fleet. A drone that fails three in a row with its own fault signature is quarantined until someone looks.
An agent can order a checkpoint onto the fleet: verified by SHA-256, only .safetensors and .gguf — pickle formats execute code — from allowlisted domains, with the owner's consent and a disk quota that never evicts their own models.
Agents record what worked and what quietly did not, and Queen attaches the machine profile automatically. A mistake made once on a 8 GB card is not repeated by every agent after.
The agent finishes its current lease, stops taking new work, updates itself and comes back — signed, silent, with a fleet version map and a kill-switch on the rollout.
A drone usually runs on a friend's or a colleague's machine. Every security decision in Hive starts from that: a leaked key must not turn their PC into someone's proxy.
ComfyUI listens on 127.0.0.1 and nothing else. The drone dials the coordinator over WSS, so it works behind any NAT or CGNAT with zero router configuration — and there is nothing to scan from outside.
Custom nodes can run arbitrary Python, so node classes are allowlisted, unknown nodes and absolute paths are refused, and every patron key carries its own quota.
The ComfyUI distribution, the node manifest and the agent release are signed with a key that does not live on the server. Compromising the coordinator must not mean code execution across the fleet.
Renders and input assets move through presigned object storage with zero egress cost and lifecycle deletion after delivery. The coordinator stays a small, cheap, boring box.
Every token is opaque and stored only as a SHA-256 hash. Drones enroll with one-time join codes and can only ever see their own leases. Enrollment, revocation, dispatch and model delivery are all in the audit log.
No sleep during a render, a health check that restarts a hung ComfyUI, a refusal to work inside OneDrive folders, and work taken only when the owner is away and the GPU is free.
The alpha is small on purpose: ten studios, each with a real production load, each talking to us directly. The price you enter on is the price you keep.
For a studio that wants to see the difference on its own footage.
For a team that already renders every day and counts hours, not experiments.
Your own coordinator on your own infrastructure — nothing of yours leaves your perimeter.
Hive has been carrying a real production load inside the studio since July 2026. Alpha means opening the same system to teams outside it — carefully, and in that order.
In production inside the studio: coordinator, fleet, agents through MCP, model delivery, silent updates.
Closed alpha: the first outside studios, each onboarded by hand, each with a direct line to the team that writes the code.
Open alpha: self-service enrollment, public agent manual, the colony journal shared across all fleets.
Beta: coordinators you run yourself, marketplace of node profiles, paid capacity between studios.
No, and you cannot. The drone establishes an outgoing secure connection to the coordinator; ComfyUI stays bound to 127.0.0.1. It works behind NAT and CGNAT with no router configuration at all.
A Windows or macOS machine with a modern GPU. The alpha covers NVIDIA and CPU on Windows and Apple Silicon on macOS; the drone installs and maintains ComfyUI itself, so nothing has to be set up by hand. Around 80 GB of free disk is recommended for models.
Yes. The panel has forms for submitting work, a live queue and a gallery. Agents are what the system was designed around, but they are not required to get value from it.
Artifacts move through presigned links straight between the drone and your storage, and are deleted by lifecycle rules after delivery. On the Foundry plan the storage is yours, so nothing you produce ever sits on our infrastructure.
The drone stops taking new work as soon as the machine is in use, finishes what it holds, and the queue moves on without it. Idle means the user is idle and the GPU is free — both, not either.
Tell us what you render and how much of it is waiting in a queue right now. If Hive fits, we will onboard you by hand and watch the first week with you.