Providers

observe(client, provider, options?)

Wraps an LLM client and returns a traced version. The returned client keeps the same API as the original.

TypeScript

const traced = observe(client, Provider.OpenAI);
const traced = observe(client, Provider.Anthropic);
const traced = observe(client, Provider.OpenRouter);

Python

traced = observe(client, Provider.OPENAI)
traced = observe(client, Provider.ANTHROPIC)
traced = observe(client, Provider.OPENROUTER)

What gets patched:

• OpenAI and OpenRouter patch client.chat.completions.create.
• Anthropic patches client.messages.create.

All other client methods remain untouched.

OpenAI

TypeScript

import OpenAI from "openai";
import { observe, Provider } from "@pulse/sdk";

const client = observe(
  new OpenAI({ apiKey: "sk-..." }),
  Provider.OpenAI
);

const res = await client.chat.completions.create({
  model: "gpt-4o",
  messages: [{ role: "user", content: "Hello" }],
});

const stream = await client.chat.completions.create({
  model: "gpt-4o",
  messages: [{ role: "user", content: "Hello" }],
  stream: true,
});

Python

from openai import OpenAI
from pulse_sdk import observe, Provider

client = observe(
  OpenAI(api_key="sk-..."),
  Provider.OPENAI
)

res = client.chat.completions.create(
  model="gpt-4o",
  messages=[{"role": "user", "content": "Hello"}],
)

Both streaming and non-streaming calls are traced. For streams, the trace is recorded once the stream completes.

Anthropic

TypeScript

import Anthropic from "@anthropic-ai/sdk";
import { observe, Provider } from "@pulse/sdk";

const client = observe(
  new Anthropic({ apiKey: "sk-ant-..." }),
  Provider.Anthropic
);

const res = await client.messages.create({
  model: "claude-3-5-sonnet-20241022",
  max_tokens: 1024,
  messages: [{ role: "user", content: "Hello" }],
});

Python

from anthropic import Anthropic
from pulse_sdk import observe, Provider

client = observe(
  Anthropic(api_key="sk-ant-..."),
  Provider.ANTHROPIC
)

res = client.messages.create(
  model="claude-3-5-sonnet-20241022",
  max_tokens=1024,
  messages=[{"role": "user", "content": "Hello"}],
)

Anthropic stop reasons are normalized: end_turn -> stop, max_tokens -> length, stop_sequence -> stop, tool_use -> tool_calls.

OpenRouter

OpenRouter uses the OpenAI client library. Pass Provider.OpenRouter so Pulse stores the correct provider and OpenRouter cost fields.

TypeScript

import OpenAI from "openai";
import { observe, Provider } from "@pulse/sdk";

const client = observe(
  new OpenAI({
    apiKey: "sk-or-...",
    baseURL: "https://openrouter.ai/api/v1",
  }),
  Provider.OpenRouter
);

Python

from openai import OpenAI
from pulse_sdk import observe, Provider

client = observe(
  OpenAI(
    api_key="sk-or-...",
    base_url="https://openrouter.ai/api/v1",
  ),
  Provider.OPENROUTER
)

When OpenRouter includes a cost field in the response, Pulse uses that value directly.

Pricing

The SDK calculates cost automatically for known models.

Model	Input	Output
gpt-4o	$2.50 / 1M	$10.00 / 1M
gpt-4o-mini	$0.15 / 1M	$0.60 / 1M
gpt-4-turbo	$10.00 / 1M	$30.00 / 1M
gpt-3.5-turbo	$0.50 / 1M	$1.50 / 1M
claude-3-5-sonnet-20241022	$3.00 / 1M	$15.00 / 1M
claude-3-5-haiku-20241022	$0.80 / 1M	$4.00 / 1M
claude-3-opus-20240229	$15.00 / 1M	$75.00 / 1M

Model aliases such as gpt-4o-2024-11-20 resolve to base model pricing. Unknown models report null cost.

Error handling

If an LLM call throws, the SDK captures an error trace and re-throws the original error.

If trace sending fails (network error or service unavailable), the SDK logs a warning and continues. Tracing never breaks application behavior.