> ## Documentation Index
> Fetch the complete documentation index at: https://docs.infercrane.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Start with an API, add your own GPUs later

> Keep one application model alias while InferCrane governs external and self-hosted inference behind it.

# Start with an API, add your own GPUs later

You do not need GPU infrastructure to get value from InferCrane. Start with a model API, add
budgets and evidence, then introduce customer-operated capacity when measurements justify it. The
application continues using one model name throughout.

## 1. Connect the API

<Warning>
  Do not create this connection until the provider, data-processing terms, residency, credential
  reference, current pricing/currency, and maximum spend envelope are approved. Connecting does not
  route a request or bill inference, but an unapproved user must stop here. InferCrane does not import
  provider invoices today, so missing billing evidence remains `unavailable` rather than an estimated
  savings claim.
</Warning>

```bash theme={"theme":"css-variables"}
export OPENROUTER_API_KEY='...'
infercrane provider connect openrouter-main \
  --model openai/gpt-4.1-mini \
  --from-env OPENROUTER_API_KEY
```

`--from-env` stores only the variable name. The variable and its value must exist in the control-plane
environment. InferCrane does not put the provider key in endpoint configuration, browser state, or
CLI output. This is non-billable configuration; it does not route a request.

## 2. Create the stable application endpoint

```bash theme={"theme":"css-variables"}
infercrane logical-model create support
infercrane endpoint create support-production \
  --model support \
  --environment production

infercrane endpoint bind support-production \
  --name api-primary \
  --connection openrouter-main \
  --request-limit 10000 \
  --cost-limit-usd 100.00 \
  --max-request-cost-usd 0.25 \
  --acknowledge-external-data \
  --enable-external

infercrane endpoint plan support-production \
  --policy manual \
  --bindings api-primary
```

Every external binding has its own privacy acknowledgement and hard authorization ceiling. The
connection itself cannot silently authorize another endpoint to spend.

## 3. Call one endpoint

<CodeGroup>
  ```python Python theme={"theme":"css-variables"}
  from openai import OpenAI

  client = OpenAI(
      base_url="https://inference.example.com/v1",
      api_key="INFERCRANE_ENDPOINT_TOKEN",
  )

  response = client.responses.create(
      model="support-production",
      input="Summarize the incident report.",
  )
  ```

  ```typescript TypeScript theme={"theme":"css-variables"}
  import OpenAI from "openai";

  const client = new OpenAI({
    baseURL: "https://inference.example.com/v1",
    apiKey: "INFERCRANE_ENDPOINT_TOKEN",
  });

  const response = await client.responses.create({
    model: "support-production",
    input: "Summarize the incident report.",
  });
  ```

  ```bash curl theme={"theme":"css-variables"}
  curl https://inference.example.com/v1/responses \
    -H "Authorization: Bearer $INFERCRANE_ENDPOINT_TOKEN" \
    -H "Content-Type: application/json" \
    -d '{"model":"support-production","input":"Summarize the incident report."}'
  ```
</CodeGroup>

## 5. Add self-hosted capacity without changing the app

```bash theme={"theme":"css-variables"}
infercrane deploy meta-llama/Llama-3.1-8B-Instruct \
  --name support-self-hosted \
  --cloud YOUR_CLOUD \
  --gpu YOUR_GPU

infercrane endpoint bind support-production \
  --name self-hosted \
  --deployment support-self-hosted

infercrane endpoint plan support-production \
  --policy primary-fallback \
  --bindings self-hosted,api-primary
```

The new plan is a candidate. Validate it with benchmark or replay evidence, evaluate Release Guard,
and promote explicitly. InferCrane does not shadow or duplicate user traffic silently.

```bash theme={"theme":"css-variables"}
infercrane endpoint guard support-production --evaluate
infercrane endpoint inspect support-production
```

The application still calls `model="support-production"`. What changed is the governed serving
plan, not the application contract.

## What this proves

* API-first onboarding without building GPU infrastructure first.
* One stable model alias across external and customer-operated compute.
* Provider keys remain server-side references.
* External data transmission and spend are explicit and bounded.
* Migration is an immutable candidate plus evidence, not a risky cutover script.

Real provider protocol behavior and billing remain provider-specific qualification. Local fixtures
prove InferCrane's policy and lifecycle logic, not a provider's production semantics.