Build, connect, and operate production inference

Start from a model, a custom OCI workload, or an endpoint you already run. InferCrane can own the complete deployment lifecycle or add policy and evidence around existing infrastructure. Your application keeps one OpenAI-compatible model name while providers, runtimes, and capacity plans change behind it. Qwen, Llama, Gemma, Kimi, and private fine-tunes all enter through the same model-artifact contract. Support is evidence-based per model, runtime, accelerator, and provider combination—not a blanket “every model everywhere” promise.

Deploy from a model

Turn a model and serving plan into a durable endpoint, then evolve it through immutable revisions.

Create an inference project

Scaffold, validate, build, and deploy a reproducible workload from one project directory.

Connect without migrating

Add health, request evidence, and deterministic diagnosis around an existing vLLM, LiteLLM, or OpenAI-compatible endpoint.

Compose your stack

Combine LiteLLM, governed OpenRouter overflow, endpoint-scoped agent sandboxes, and signed training artifacts without merging ownership boundaries.

Ship a safer revision

Combine bounded performance validation with signed task-quality evidence and keep production unchanged when policy rejects it.

Handle uncertain capacity

Resume long operations, explain cold starts, compare measured serving plans, and govern external overflow.

One stable application contract

from openai import OpenAI

client = OpenAI(
    base_url="https://infercrane.example/v1",
    api_key="${INFERCRANE_API_KEY}",
)

response = client.responses.create(
    model="coder-production",
    input="Review this function for concurrency bugs.",
)

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://infercrane.example/v1",
  apiKey: process.env.INFERCRANE_API_KEY,
});

const response = await client.responses.create({
  model: "coder-production",
  input: "Review this function for concurrency bugs.",
});

curl https://infercrane.example/v1/responses \
  -H "Authorization: Bearer $INFERCRANE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"coder-production","input":"Review this function for concurrency bugs."}'

coder-production is the stable product identity. A binding may resolve it to an InferCrane-managed deployment, an adopted workload, or governed external capacity. Applications do not need provider credentials or physical model names when the serving plan changes.

Choose how much InferCrane owns

Start here	InferCrane can observe	InferCrane can route	InferCrane can create or delete compute
`observe-only` connection	Yes	No	No
`traffic-managed` connection	Yes	Yes	No
lifecycle-managed deployment	Yes	Yes	Yes

This ownership ladder lets a platform team prove value before handing over traffic or provider lifecycle. Promotion is explicit; connecting an endpoint never silently changes production routing.

The proof loop

InferCrane does not use an LLM to decide production promotion. Policy inputs, measurements, reason codes, and revision identities are persisted so the result can be inspected and reproduced.

Examples use qualified product surfaces, but infrastructure and runtime combinations have independent evidence states. Check the capability matrix before relying on a combination in production.

Run the proof loop locally

git clone https://github.com/infercrane/infercrane.git
cd infercrane
make demo

The hermetic demo exercises adoption, a real proxied fixture request, Request Inspector, Doctor, immutable candidate creation, Release Guard rejection, deterministic explanation, and cleanup. It does not provision a GPU and cannot be used as provider or performance evidence.

GET STARTED

BUILD

OPERATE

INTELLIGENCE

INTEGRATE

REFERENCE

PROJECT

What you can build

Build, connect, and operate production inference

Deploy from a model

Create an inference project

Connect without migrating

Compose your stack

Ship a safer revision

Handle uncertain capacity

One stable application contract

Choose how much InferCrane owns

The proof loop

Run the proof loop locally

​Build, connect, and operate production inference

Deploy from a model

Create an inference project

Connect without migrating

Compose your stack

Ship a safer revision

Handle uncertain capacity

​One stable application contract

​Choose how much InferCrane owns

​The proof loop

​Run the proof loop locally

Build, connect, and operate production inference

One stable application contract

Choose how much InferCrane owns

The proof loop

Run the proof loop locally