ClusterFlock

AI compute is fragmented by default. A team accumulates a DGX Spark here, a couple of RTX workstations there, a few Macs with Apple Silicon, maybe Jetson units at the edge, and each sits in its own silo: underused, and impossible to address as one resource.

Renting equivalent capacity in the cloud is expensive and sends your data off-premises. The hardware to run serious inference is often already in the building; there was no way to make it behave as a single machine.

ClusterFlock is a distributed AI compute orchestrator that unifies heterogeneous GPU hardware into one cluster behind a central hub called nCore. It automatically detects and VRAM-profiles every node, then intelligently bin-packs models across the fleet: no manual configuration, self-adapting and self-healing even while a mission is running.

It exposes an OpenAI-compatible API on port 1919, so any OpenAI SDK, LangChain, or LiteLLM app points at it and works unchanged. Three routing modes cover the spectrum: FANOUT broadcasts to all nodes and synthesizes the best answer, SPEED picks the fastest single endpoint, MANUAL lets you choose the model.

You describe a goal. ClusterFlock spins up sandboxed containers, a showrunner LLM coordinates a flock of worker LLMs, dispatches tasks, evaluates results, and iterates until the job is done: a genuine mixture-of-agents architecture running on your own metal.

Real-time telemetry streams GPU utilization, VRAM, model status, and tokens per second across every node, every five seconds. Four agent types cover the hardware landscape (Spark for DGX and GB10, Linux for CUDA, Mac for Apple Silicon, LMS for LM Studio) across llama.cpp, LM Studio, Metal, and CUDA backends.

ClusterFlock is proof that Notum builds infrastructure, not just applications: the layer other AI products run on top of. It stakes out a piece of the AI-compute-orchestration stack, and it is open source under the MIT license precisely to seed adoption. The code is the claim.

ClusterFlock ships in the open: MIT-licensed, currently at v0.7.10, and public on GitHub. That is the rarest kind of proof in deep tech: not a claim about capability, but running code anyone can inspect, clone, and benchmark.