The agent-operable control plane for bare-metal fleets.
Turn racks of bare metal, across any hardware, provider and site, into ready-to-use Kubernetes and Slurm clusters, and keep them that way. Your engineers work on what runs inside the cluster, not on keeping the cluster running. Open source, and built to be run by a lean team and, increasingly, an AI operator.
- Kubernetes and Slurm clusters on your own hardware, installed, upgraded and healed for you
- From power-on to production with no PXE, no kickstart and no snowflakes
- One encrypted mesh across every site and provider
- Every action typed, health-gated and reversible
$ curl -fsSL get.durantic.io | sh # installer ships with the open-source launch
INFO node-01 power on, agent online
INFO node-01 provisioned, kexec into installed OS
INFO node-01 joined mesh, 4 peers
INFO node-01 config release verified and converged, READY
Big fleets still run on people and scripts.
A startup with twenty servers, or everything in one cloud, doesn't feel this. It starts at a few hundred machines, with mixed hardware generations, several sites and providers, and GPU nodes failing every few hours. Teams stitch together five to eight tools that don't agree on what a machine is, and the glue is the one engineer who knows the fleet. Every hour spent keeping Kubernetes or Slurm alive on that metal is an hour not spent on the models, jobs and services the cluster exists for.
- Months of plumbing before the first job runs
- Slow, fragile provisioning
- Racked machines that aren't earning
- Hand-edited network configs across sites
- Kubernetes and Slurm upgrades nobody dares to start
- Config drift and risky rollouts
- A 3am page to the one person who knows
- Operations headcount that grows with every rack
Make the fleet safe to hand over.
Every fleet gets handed to someone who wasn't there when it was built: a new hire, an on-call rotation, and soon an AI operator. Scripts, SSH sessions and tribal knowledge don't survive the handover. They only work inside the head of the person who wrote them.
Durantic gives the fleet a contract. One agent on every machine. Every change is a typed action the control plane understands: provision, rebuild, roll out, roll back. Every action is health-gated, audited and reversible, no matter who starts it.
$ ssh root@node-412 # vim /etc/netplan/50-cloud-init.yaml # ./reimage.sh --force # nobody remembers why --force # exit $ echo 'did node-412, will do the rest tomorrow' >> ops-log.txt
action: rebuild machines: role=gpu-worker, site=COLO-AMS-1 image: ubuntu-26.04-gpu gate: healthy before next batch rollback: previous image, one step cosigned: on-call engineer (2FA)
A machine's first two minutes.
- 00:00
Power on
The machine boots, finds the control plane and reports exactly what it is: GPUs, NICs, disks, firmware. No PXE server, no kickstart, no spreadsheet of MAC addresses.
- 00:40
Provision
It pulls its operating system as an OCI image and kexecs straight into it. Every machine in a role gets the same image, byte for byte.
- 01:30
Connect
It joins the encrypted mesh. Colo, cloud, edge and the rack under the desk become one network, with no peer lists to edit.
- 01:50
Configure
Its configuration is already in place: the agent applied it during install, reboot-requiring settings included, so its first boot is a production boot. From then on it verifies every bundle by digest and converges locally, with no central master.
- 02:00
Operate
It's in the fleet. From here every change is a typed, health-gated, reversible action, whether an engineer or the reeve starts it.
Open-source core. Hosted control plane. AI operator on top.
Durantic Core
One agent per machine. Unattended provisioning from power-on, an encrypted WireGuard + BGP mesh across every site, signed configuration releases and official Kubernetes and Slurm bundles. Self-host it free.
→ GitHubDurantic Cloud
A hosted control plane for teams that don't want to run one: fleet view, API and CLI, with RBAC and audit. Your machines and your data plane stay yours.
→ PricingThe reeve
An AI operator that sees the whole fleet: rollouts, rebuilds, health. Your engineers talk to it in the control plane. It diagnoses issues and proposes fixes, and anything privileged, like terminal access or a rebuild, needs a human to co-sign with 2FA.
Built for the fleets cloud tooling forgot.
Most infrastructure tools assume one homogeneous cloud. Durantic was designed for the opposite: mixed hardware, fragmented capacity, several sites and physical machines.
Heterogeneous by design
Mixed hardware generations, multiple providers, colocated racks: one operational model.
Bare metal-native
Built for physical machines and firmware-level control, not bolted onto cloud assumptions.
Clusters, not just machines
Durantic delivers working Kubernetes and Slurm clusters: control plane, workers, networking, upgrades and node replacement, all handled. Your team works inside the cluster, not on it. Bring your own stack if you prefer.
Safe to automate
Every action is typed, health-gated, audited and reversible. That is what makes an AI operator safe to run.
Open source
Run it, read it, extend it. No lock-in on the core.
An AI operator that earns its autonomy.
- STEP 1
Observes and advises
Reads the fleet, explains what it sees, proposes the fix.
- STEP 2
Acts with your 2FA co-sign
Runs the typed action once an engineer approves it with two-factor sign-off.
- STEP 3
Handles delegated work
Takes the classes of work you delegate, with every action audited.
The reeve works through the same typed, health-gated actions your engineers use. Every fix it makes, with the customer's consent, improves the library of actions it uses on the next fleet.
Bring your own model: we run inference in Durantic Cloud, you can use your own API key, and air-gapped deployments can plug in their own LLM.
Fleets of 50 to 5,000 machines, run by 1 to 15 engineers.
GPU clouds and neoclouds
Bring racks into service unattended as Kubernetes or Slurm capacity, keep mixed GPU generations on one operational model, and stop hand-building every new site.
Bare-metal hosts and colocation
Run customer-facing bare metal across facilities with one agent and one mesh. With Equinix Metal shut down, more of this market runs its own machine layer.
Quant and trading firms
Own data centres, strict change control and deterministic hardware: versioned, digest-pinned releases and reversible actions fit how you already work.
Sovereign, private and air-gapped AI
The cluster experience teams expect from a public cloud, on hardware you own, in facilities you choose. Self-host the open-source core, with your own model behind the reeve.
Research and HPC
Hand researchers Slurm or Kubernetes clusters that install, upgrade and heal themselves. Your team spends its time on the science and the jobs, not on the cluster.
Install. Pilot. Scale.
- 01
Install
Run Durantic Core on a spare rack. Your first machine is provisioned in minutes.
- 02
Pilot
A 30–60 day paid pilot on your fleet, with a success metric agreed up front.
- 03
Scale
Add sites and machines, choose self-hosted with support or Durantic Cloud, and turn on the reeve.