GPUField project directory / 01

HuxKan

An independent control-plane project for self-hosted GPU fleets, with a gateway agent that stays beside the model runtime.

Preview software. Review the current implementation limits before using it with untrusted workloads.

CONTROL LOOP / TWO WAYKAN ↔ HUX

Kan can run in a cloud environment or inside a customer-controlled on-prem location. Desired state moves toward Hux; operational signals return to Kan.

One control plane / many GPU locations

Hux lives with the GPUs.
Kan sees the fleet.

Kan can run in a cloud or customer on-prem location. Each map marker is an illustrative deployment region—not a claimed customer or integration. Hux stays close to each GPU runtime and maintains a two-way operational channel with Kan.

KAN → HUX policy · desired state · approved commandsHUX → KAN health · inventory · telemetry · metering

Coastlines: Natural Earth 1:110m, public domain · Equirectangular projection · No political borders · Markers are illustrative

01 / PRIVATECustomer datacenters

Kan may run on-prem, while Hux stays beside GPU workers inside operator-controlled infrastructure.

02 / GPU CLOUDS — EXAMPLESNeysa · Yotta · E2E · CoreWeave · Nebius · Firmus

Illustrative places a fleet could run. No partnership or current integration is implied.

03 / HYPERSCALERS — EXAMPLESAWS · Google Cloud · Azure

Kan or GPU capacity may run in cloud environments while preserving the same agent boundary.

Boundary first

Inference stays on the data plane. The control plane exchanges configuration, health, inventory, telemetry, and metering—not prompts or completions.

Current shape

Built around a bidirectional agent connection.

Organize

Model infrastructure is represented as organizations, fleets, nodes, and projects.

Control

Kan sends desired routing state, policies, access configuration, and approved node operations toward Hux.

Observe

Hux returns health, hardware inventory, telemetry, request metering, and audit events to Kan.

Implementation note

What the current build does not prove

It does not establish production-grade multi-tenant security solely by requiring login.
It does not establish fleet-global quota consistency from node-local counters.
It has not completed qualification for every GPU, runtime, cloud, or failure mode.