Skip to main content
1

Install the SDK

pip install pikarc
2

Get your API key

Sign up at app.pikarc.dev and copy your API key from Settings > API Key. The key format is lg_<prefix>_<secret>.
3

Wrap your model call

from pikarc import AsyncPikarc, PikarcBlockedError
from openai import AsyncOpenAI

openai = AsyncOpenAI()
guard = AsyncPikarc(
    api_key="lg_xxxx_your_api_key",
    base_url="http://localhost:8000",
)

async def chat(user_id: str, message: str) -> str:
    try:
        async with guard.run(user_id=user_id) as run:
            response = await run.model_call(
                fn=lambda: openai.chat.completions.create(
                    model="gpt-4o",
                    messages=[{"role": "user", "content": message}],
                ),
                model="gpt-4o",
                input_data={"messages": [{"role": "user", "content": message}]},
                token_extractor=lambda r: (
                    r.usage.prompt_tokens,
                    r.usage.completion_tokens,
                ),
            )
            return response.choices[0].message.content or ""

    except PikarcBlockedError as e:
        return f"Blocked: {e.reason}"

    finally:
        await guard.close()
4

See it in the dashboard

Open your dashboard to see the run, step timeline, token usage, and cost breakdown.

​
What just happened?

  1. guard.run() called POST /v1/runs/ — Pikarc checked the kill switch, budgets, and concurrency limits. The run was ALLOWED.
  2. run.model_call() called POST /v1/runs/{id}/steps — Pikarc evaluated guardrails again before the model call.
  3. After OpenAI responded, the SDK called PATCH /v1/runs/{id}/steps/{step_id} to report token usage and duration.
  4. When the async with block exited, the SDK called POST /v1/runs/{id}/end to mark the run as COMPLETED.
If any guardrail check had failed (budget exceeded, kill switch active, etc.), PikarcBlockedError would have been raised before the model call executed.

​
Next steps