Skip to main content
Integrations record what only the framework knows about, such as agent steps, tool calls and sessions, and add relevant lessons to model requests. Install the one for your framework next to the SDK; auto_instrument() / autoInstrument() still records model calls over HTTP and MCP calls.

LangChain and LangGraph

Middleware for create_agent, or a chat-model wrapper for LangGraph and direct model use.

OpenAI Agents SDK

Function tools and agent tools of your agents.

Claude Code

Each session as a trace, with lessons added to every prompt.

Pi

Each session as a trace, with lessons added to Pi’s context.
Model calls that an integration records itself (the LangChain middleware and chat-model wrapper, Claude Code and Pi) are stored as conversations, not HTTP requests, so Rerun is not available for them. Model calls captured over HTTP by auto_instrument() can be rerun.

Building an integration

Integrations use only the public SDK: the step API (start_step, record_step, retrieve_lessons, …), the cross-process functions, and sovara.integration / @sovara/sdk/integration. Those helpers never let a Sovara outage break the host, retry only briefly, and convert arbitrary values to JSON safely. See the Python and TypeScript API references.