integrations.mdx

title	Framework integrations
description	Wrap any agent framework so Volo captures every model + tool call — six frameworks, one wrap() shape.

Volo instruments any framework through a thin adapter: wrap() swaps the framework's model/tool objects with Volo capture proxies, and import_*_otel() turns an OTel trace into a Volo Recording. The shape is identical across frameworks, so learning one is learning all six.

Framework	Package	wrap() swaps
LangGraph	volo-langgraph	graph model + tool nodes
OpenAI Agents SDK	volo-openai-agents	agent.model + tools
CrewAI	volo-crewai	each agent's llm + tools; kickoff decision
AutoGen	volo-autogen	model_client (v0.4) or llm (0.2) + tools
Pydantic AI	volo-pydantic-ai	agent.model + tools
Semantic Kernel	volo-semantic-kernel	kernel.services + plugins

Usage

from volo_autogen import wrap
from volo_sdk import record

agent = wrap(my_autogen_agent, model=my_model, tools=my_tools)

with record(agent_name="support-bot", framework="autogen"):
    agent.run()          # every model + tool call is captured
# → ./.volo/recordings/<run_id>.json — now replay, score, fuzz, migrate, red-team it

Each wrap() also records a decision step per run so the trajectory keeps the framework's structure visible. Prefer to instrument via OpenTelemetry instead? Every integration ships an importer:

from volo_pydantic_ai import import_pydantic_ai_otel
recording = import_pydantic_ai_otel("trace.jsonl", agent_name="support-bot")

The importers accept JSONL, OTLP-JSON, or a bare span array (the shared import_otel_trace seam), so a trace exported from any OTel-instrumented run becomes a replayable recording — no code change to the agent.

Once captured, everything applies

A recording is a recording regardless of which framework produced it, so the whole platform works on all six: deterministic replay & scoring, MCP simulation, red-teaming, the migration lab, personas, and the long-horizon rig.