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

# SDK Integration

> Endpoints and parameter mappings for the OpenAI and Anthropic SDKs.

This is the reference for integrating the OpenAI and Anthropic SDKs with SERV. For a five-minute setup, start with the [Quickstart](./introduction).

SERV exposes three HTTP endpoints under one base URL: `https://inference-api.openserv.ai`

| Endpoint                    | Shape     | Use it for                                                                          |
| --------------------------- | --------- | ----------------------------------------------------------------------------------- |
| `POST /v1/chat/completions` | OpenAI    | Universal. Works with every model in the catalog.                                   |
| `POST /v1/responses`        | OpenAI    | OpenAI models, with streamed reasoning summaries.                                   |
| `POST /v1/messages`         | Anthropic | Claude and most other providers. See [endpoint compatibility](./api/compatibility). |

## OpenAI SDK

### Chat completions

```js theme={null}
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://inference-api.openserv.ai/v1",
  apiKey: process.env.SERV_API_KEY,
});

const completion = await client.chat.completions.create({
  model: "gpt-5.4-mini",
  messages: [
    { role: "system", content: "You are a concise assistant." },
    { role: "user", content: "What is a CPU register?" },
  ],
});

console.log(completion.choices[0].message.content);
```

### Responses

Use the Responses API to receive the reasoning trace alongside the answer. OpenAI models only.

```js theme={null}
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://inference-api.openserv.ai/v1",
  apiKey: process.env.SERV_API_KEY,
});

const response = await client.responses.create({
  model: "gpt-5.4",
  instructions: "You are a careful reasoner.",
  input: "What is the integral of x^2 from 0 to 3?",
});

console.log(response.output_text);
```

## Anthropic SDK

### Messages

```js theme={null}
import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  baseURL: "https://inference-api.openserv.ai",
  authToken: process.env.SERV_API_KEY,
});

const message = await client.messages.create({
  model: "claude-haiku-4.5",
  max_tokens: 1024,
  system: "You answer in one sentence.",
  messages: [{ role: "user", content: "What is a CPU register?" }],
});

console.log(message.content.find(b => b.type === "text").text);
```

## Important details

### Base URL differs by SDK

| SDK           | Base URL                               | Why                                                                                                                  |
| ------------- | -------------------------------------- | -------------------------------------------------------------------------------------------------------------------- |
| OpenAI SDK    | `https://inference-api.openserv.ai/v1` | The OpenAI SDK expects `/v1` in the base URL.                                                                        |
| Anthropic SDK | `https://inference-api.openserv.ai`    | The Anthropic SDK appends `/v1/messages` itself. Including `/v1` yourself would call `/v1/v1/messages`, which fails. |

### Auth field differs by SDK

| SDK           | Field       | Notes                                                                                                          |
| ------------- | ----------- | -------------------------------------------------------------------------------------------------------------- |
| OpenAI SDK    | `apiKey`    | Standard.                                                                                                      |
| Anthropic SDK | `authToken` | `apiKey` also works, but `authToken` keeps `ANTHROPIC_API_KEY` free if you ever fall back to direct Anthropic. |

### A system prompt is required

Every request needs a system, developer, or instructions message. Requests without one are rejected:

```
A system prompt is required. Please include a system or developer message in your request.
```

Where the system prompt goes depends on the endpoint:

| Endpoint               | Where the system prompt goes                   |
| ---------------------- | ---------------------------------------------- |
| `/v1/chat/completions` | a `{ role: "system", content: "..." }` message |
| `/v1/responses`        | top-level `instructions`                       |
| `/v1/messages`         | top-level `system`                             |

## Parameter map

Moving an integration across SDKs comes down to this mapping.

| Concept                           | OpenAI Chat                                                 | OpenAI Responses                                    | Anthropic Messages                                                   |
| --------------------------------- | ----------------------------------------------------------- | --------------------------------------------------- | -------------------------------------------------------------------- |
| HTTP path                         | `/v1/chat/completions`                                      | `/v1/responses`                                     | `/v1/messages`                                                       |
| Auth field (SDK constructor)      | `apiKey`                                                    | `apiKey`                                            | `authToken`                                                          |
| `baseURL` suffix to use with SERV | `/v1`                                                       | `/v1`                                               | *(none)*                                                             |
| Token cap field                   | `max_completion_tokens`                                     | `max_output_tokens`                                 | `max_tokens` *(required)*                                            |
| System prompt                     | message with `role:"system"`                                | top-level `instructions`                            | top-level `system`                                                   |
| User message shape                | `{role, content}` in `messages[]`                           | top-level `input` (string or array)                 | `{role, content}` in `messages[]`                                    |
| Reasoning-effort control          | `reasoning_effort`                                          | `reasoning: { effort, summary }`                    | `thinking: { type:"enabled", budget_tokens }`                        |
| Streaming                         | `stream: true`                                              | `stream: true`                                      | `stream: true`                                                       |
| Stop sequences                    | `stop`                                                      | n/a                                                 | `stop_sequences`                                                     |
| Tool schema                       | `tools: [{type:"function", function:{name, parameters}}]`   | `tools: [{type:"function", ...}]`                   | `tools: [{name, input_schema}]` *(no nested `function:`)*            |
| Tool choice                       | `tool_choice: "auto" \| {type:"function", function:{name}}` | `tool_choice: ...`                                  | `tool_choice: "auto" \| "any" \| {type:"tool", name}`                |
| Response text                     | `choices[0].message.content`                                | `output_text` or `output[]` blocks                  | `content[]` array, find the `type === "text"` block                  |
| Token usage                       | `usage.prompt_tokens / completion_tokens / total_tokens`    | `usage.input_tokens / output_tokens / total_tokens` | `usage.input_tokens / output_tokens`                                 |
| Cache metrics                     | `usage.prompt_tokens_details.cached_tokens`                 | `usage.input_tokens_details.cached_tokens`          | `usage.cache_read_input_tokens`, `usage.cache_creation_input_tokens` |

## See also

* [SERV Tools](./tools) — enable the prompt-injection guard and shadow agents via tool definitions.
* [Endpoint compatibility](./api/compatibility) — which models work on which endpoint.
* [Quickstart](./introduction) — the five-minute setup.
* [SDK Migration](./sdk-migration) — Python, Vercel AI SDK, LangChain, and raw `fetch`.
* [Models](./models) — pricing and context windows.