Pydantic AI

About the Pydantic AI integration

Pydantic AI is OpenTelemetry-native. Agent.instrument_all() or logfire.instrument_pydantic_ai() emits GenAI semantic-convention spans that telemetry.dev ingests natively. Pydantic AI emits gen_ai.usage.* attributes on model spans, and telemetry.dev normalizes them into model, provider, token, latency, and cost fields.

Key features

• Agent runs and model calls in one trace: Follow each run across model requests, tools, and responses.
• Native Pydantic AI usage attributes: Ingest gen_ai.usage.* from model spans without custom mapping.
• Server-side cost from real token usage: Cost is computed from emitted token counts and current model pricing.
• Standard OTLP/HTTP: Keep Pydantic AI's portable OpenTelemetry instrumentation without a proprietary tracing SDK.
• Works with or without Logfire: Enable spans with Agent.instrument_all() directly or logfire.instrument_pydantic_ai() in an existing Logfire setup.

Get started

1. Install Pydantic AI and the OpenTelemetry OTLP/HTTP exporter:

   pip install pydantic-ai opentelemetry-sdk opentelemetry-exporter-otlp-proto-http

2. Create a telemetry.dev project API key and configure the exporter:

   export TELEMETRY_DEV_API_KEY="td_live_..."
   export OTEL_EXPORTER_OTLP_ENDPOINT="https://ingest.telemetry.dev"
   export OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer%20$TELEMETRY_DEV_API_KEY"
   export OTEL_SERVICE_NAME="your-service-name"
   export OTEL_RESOURCE_ATTRIBUTES="deployment.environment.name=production"

3. Configure raw OpenTelemetry and instrument every Pydantic AI agent:

   from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
   from opentelemetry.sdk.trace import TracerProvider
   from opentelemetry.sdk.trace.export import BatchSpanProcessor
   from opentelemetry.trace import set_tracer_provider
   from pydantic_ai import Agent
   
   provider = TracerProvider()
   provider.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))
   set_tracer_provider(provider)
   
   Agent.instrument_all()
   
   agent = Agent("openai:gpt-5.2")
   result = agent.run_sync("What is the capital of France?")
   print(result.output)

The OTLP exporter reads the endpoint and authorization header from the environment and sends traces to https://ingest.telemetry.dev/v1/traces.

What gets captured

By default, Agent.instrument_all() exports full content: prompts, responses, message history, tool arguments and results, final outputs, model request parameters, and any binary content in messages. To reduce captured content, disable message content and model request parameters:

from pydantic_ai import Agent, InstrumentationSettings

Agent.instrument_all(
    InstrumentationSettings(
        include_content=False,
        include_model_request_parameters=False,
    )
)

This is a lower-content configuration, not metadata-only: include_content=False suppresses most prompts, responses, and tool payloads, but structured-output validation errors and ModelRetry feedback can still appear in pydantic_ai.all_messages and gen_ai.input.messages. include_model_request_parameters=False omits the full model_request_parameters attribute. Even with both settings off, tool definitions (gen_ai.tool.definitions), agent descriptions, run metadata, standard gen_ai.request.* settings, and retry or validation feedback can still be exported. InstrumentationSettings(include_binary_content=False) keeps text content but drops inline binary data such as images and files.

Already exporting OTLP/HTTP?

If your app already configures an OTLP/HTTP exporter and tracer provider, keep that code and only set the environment variables from step 2. Call Agent.instrument_all() at startup, or keep logfire.instrument_pydantic_ai() if Logfire already provides your instrumentation. Pydantic AI spans will flow through your existing exporter to telemetry.dev.