Ostra Pro with OpenAI and Anthropic clients

Use Ostra Pro from the OpenAI and Anthropic SDKs, LangChain, the Vercel AI SDK and any tool with an OpenAI-compatible base URL — base URLs, keys, sessions, streaming and cold starts.

Ostra Pro speaks two API formats, so most LLM tooling works by changing three settings:

SettingOpenAI-compatible clientsAnthropic clients
Base URLhttps://ostra.testrelic.ai/api/v1https://ostra.testrelic.ai/api
API keyYour Ostra key (ostk_live_…)Your Ostra key
Modelostra-proostra-pro (any value is accepted)

Every example reads the key from OSTRA_API_KEY.

OpenAI SDK

import os

from openai import OpenAI

client = OpenAI(
    base_url="https://ostra.testrelic.ai/api/v1",
    api_key=os.environ["OSTRA_API_KEY"],
    default_headers={"x-ostra-session": "checkout-debugging"},  # one value per conversation
    max_retries=6,  # rides out a cold start on non-streaming calls
)

completion = client.chat.completions.create(
    model="ostra-pro",
    reasoning_effort="medium",
    messages=[
        {"role": "system", "content": "You are a senior test engineer."},
        {"role": "user", "content": "Why would page.waitForURL time out only on CI?"},
    ],
)
print(completion.choices[0].message.content)

To turn thinking off for a quick, cheap answer, pass extra_body={"chat_template_kwargs": {"enable_thinking": False}} in Python, or add chat_template_kwargs: { enable_thinking: false } to the request body in TypeScript.

Anthropic SDK

import os

import anthropic

client = anthropic.Anthropic(
    base_url="https://ostra.testrelic.ai/api",
    api_key=os.environ["OSTRA_API_KEY"],
    default_headers={"x-ostra-session": "checkout-debugging"},
)

message = client.messages.create(
    model="ostra-pro",
    max_tokens=2048,
    messages=[{"role": "user", "content": "Write a Playwright test that the cart badge shows 3 items."}],
)
# A thinking block can come before the answer: print the text blocks.
print("".join(block.text for block in message.content if block.type == "text"))

Anthropic-format differences: output_config.effort above medium runs as medium; tool_choice any or tool runs as auto; count_tokens isn't supported. See Create a message.

LangChain

import os

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="ostra-pro",
    base_url="https://ostra.testrelic.ai/api/v1",
    api_key=os.environ["OSTRA_API_KEY"],
)
print(llm.invoke("List edge cases for a coupon code input field.").content)

Vercel AI SDK

Use the OpenAI-compatible provider, which calls Chat Completions:

import { createOpenAICompatible } from '@ai-sdk/openai-compatible';
import { generateText } from 'ai';

const ostra = createOpenAICompatible({
  name: 'ostra',
  baseURL: 'https://ostra.testrelic.ai/api/v1',
  apiKey: process.env.OSTRA_API_KEY,
});

const { text } = await generateText({
  model: ostra('ostra-pro'),
  prompt: 'Turn this bug report into test cases: "Coupon accepted after expiry date".',
});
console.log(text);

Other tools

Any tool that lets you set an OpenAI-compatible base URL (sometimes called "custom OpenAI endpoint" or "OpenAI-compatible provider") works the same way: base URL https://ostra.testrelic.ai/api/v1, your Ostra key as the API key, model ostra-pro. Tools that only speak the newer OpenAI Responses API won't work: Ostra Pro serves Chat Completions.

Claude Code

Claude Code can run on Ostra Pro through the Anthropic-compatible endpoint:

export ANTHROPIC_BASE_URL="https://ostra.testrelic.ai/api"
export ANTHROPIC_AUTH_TOKEN="$OSTRA_API_KEY"
export ANTHROPIC_MODEL="ostra-pro"
claude

Supported through the Anthropic-compatible endpoint

Claude Code talks to Ostra Pro through /api/v1/messages, with the adaptations listed above. If a Claude Code feature doesn't work as expected, contact TestRelic support with the request id from Logs.

Good practice

  • One session per conversation. Send x-ostra-session with a stable value per conversation or agent run, so follow-up turns usually stay on one replica and more of them bill as cached input. See Sessions and caching.
  • Stream, or allow retries. The first request after 5 quiet minutes starts the model. Streaming waits for it; non-streaming calls may get 503 with Retry-After, which SDK retries absorb if you allow enough of them. See Cold starts.
  • Warm up before bursts. POST /api/v1/warm is free.
  • Size max_tokens. The most a request can cost is held from your balance before it runs; a smaller max_tokens means a smaller hold.
  • Use a dedicated key with a monthly spend cap for each tool.
Was this page helpful?

On this page