Where agents, models and jobs run.

Runtimes are isolated environments close to the machine or in the fleet. Each carries a signed policy, an identity derived from the node's hardware key, and a record a third party can check.

Live · compute fleet, silos Request access NVIDIA compatibility
A compute node serving models
01What runs
An agent confined in a silo with one signed channel

Four ways to run.

Send a job and get the result back, start a silo for an agent, or bring your own machines into the pool.

Compute

Send a heavy job (a render, a cloth drape, a simulation, a model run) and get the result back. Jobs run on the Ipnops fleet or on machines you bring; the worker calls out, so it works behind any firewall.

Supported
Silos

A Firecracker microVM started for one agent, with its own kernel and disk and no network interface. The agent is reached through the API or over MCP.

Supported
Agent runtime

One contract over every sandbox: a policy the operator signed is verified before anything runs, and decisions are committed in signed batches.

Supported
Job programs

Pinned environments with no network, read-only inputs and one scratch folder. A job gets a GPU or is refused.

Supported
Vault

Sealed data and keys: passkey-held for people, hardware-held for nodes. The vault and its silos.

Supported
02How it is held
A chip with a hardware identity and a secure element
identity

Derived from the node's key.

An agent's identity is a digest of node, agent name, policy, model and epoch, endorsed by the node's hardware key. No agent key exists to leak.

policy

Signed before it runs.

The runtime verifies the signature, issuer, window and site before any sandbox is told anything.

placement

Jobs name what they need.

GPU class, memory, cores and operating system. Machines take jobs on a lease; a machine that goes away loses it and the job moves.

record

A third party can check it.

Decisions are numbered by the node and committed in signed batches. A verified kill order tears the sandboxes down and records it.

03Across silicon
An edge AI module on its carrier board
GPU classes

Workers advertise the GPUs they hold, and jobs place on what they name. Jobs run on unified-memory machines, on CUDA workstations and on a Windows render host.

Supported
Attestation grades

A node earns soft, device, se, tee, dice or tpm from how it holds its key. A reading is graded by the weakest link in its chain.

Supported
Embedded modules

Small embedded compute modules as placement targets, with rootless containers and device access.

Available on request
Across architectures

Identity and attestation run on x86 servers, Arm single-board computers and microcontrollers, each graded by how it holds its key.

Supported

Run it where the machine is.

Tell us the models, the machines and the policy.