Integrations

Maida is framework-agnostic at its core. Adapters translate framework callbacks into the same LLM_CALL and TOOL_CALL structure; framework-specific details remain in meta, and every payload passes through Maida’s redaction and truncation before local storage.

Integration	Install or connect	Guide
LangChain / LangGraph	uv add "maida-ai[langchain]>=0.6.1"	Callback handler
OpenAI Agents SDK	uv add "maida-ai[openai]>=0.6.1"	Tracing adapter
CrewAI	unsupported after v0.5.3 (pin maida-ai[crewai]==0.5.3)	Execution-hook adapter
Langfuse	Existing completed traces	Langfuse import guide

CrewAI support was dropped after v0.5.3 because crewai dependency conflicts block Python 3.14 and openai upgrades. The adapter code remains in-tree; the [crewai] extra does not. See the CrewAI guide for details and the v0.5.3 pin.

Download the deterministic offline examples directly:

• LangChain and LangGraph example

• OpenAI Agents example

• CrewAI example (historical; requires installing crewai yourself on post-v0.5.3 checkouts)

Shared contract

• Framework packages are optional and loaded only when their adapter is imported.

• Equivalent model and tool activity produces framework-neutral trace events.

• Adapters require an active @trace or traced_run(...); they never create unrelated runs.

• Framework errors remain the application’s errors and are recorded as error-status activity.

• Prompts, responses, arguments, results, errors, and metadata use the same recursive redaction and byte limits as direct SDK calls.

If an optional dependency is absent, importing its adapter fails immediately with an ImportError that names the required Maida extra. Core Maida remains importable without any framework installed.