Skip to main content
Open-source inference control plane
Start the quickstart View on GitHub
Build or connectSafe releasesExplainable operationsNo Kubernetes required
Deploy a model
Send a streaming request
Build a deployment from a model or custom OCI workload, or connect inference you already run. InferCrane gives applications one OpenAI-compatible logical endpoint while durable control-plane operations provision, reconcile, scale, and safely replace workers behind it. Providers and runtimes integrate through explicit contracts; Kubernetes is not required.
requestOpenAI client
stable identityendpoint
policyroute generation
healthyruntime replicas

​
The operational layer around inference runtimes

Build and deploy

Start from a model, recipe, or immutable custom OCI workload. Disconnecting the CLI does not abandon the operation.

Protect releases

Compare active and candidate revisions with deterministic, persisted Release Guard policy.

Understand behavior

Explain scaling, rollout, degradation, and cold starts from recorded state and measurements.

​
Start from where you are

I have a model

Plan a qualified serving configuration and create a durable endpoint.

I have an endpoint

Add health and operational evidence without migrating or transferring lifecycle ownership.

I am evaluating

Run the complete GPU-free control and request path locally in under five minutes.

​
One product, two compute modes

A registered elastic backend provisions runtime replicas. InferCrane reconciles desired capacity, routes only to ready workers, and fences scale-down before provider termination.
InferCrane is provider- and runtime-extensible by design. Provider and runtime capabilities are declared and conformance-tested; each adapter still has an independent qualification status. vLLM, SGLang, and custom OCI have separate evidence states, as do RunPod and narrow AWS BYOC adapters. Consult capability status before relying on a combination.

​
Evidence, not guesses

InferCrane records durable operation events, immutable model and revision identity, normalized request telemetry, benchmark reproduction metadata, and deterministic policy decisions. It does not fabricate provider pricing or timing boundaries that the provider does not expose.

Five-minute local tour

Run the GPU-free development stack and send your first OpenAI-compatible request.

Read the architecture

See the control-plane and data-plane boundaries behind durable operations.

See production patterns

Deploy a model, connect LiteLLM, scope a sandbox, attach trained artifact lineage, and reject a bad release.