- Conversations — instrument agent turns from your application code.
- MCP — wrap an MCP server so every call it handles is captured for you.
Which one do I use
Conversations SDK
Reach for this when your agent is application code that produces a reply per turn. You control when a turn begins and ends, and you attach the user, the input, and the output.
MCP wrapper
Reach for this when your agent is an MCP server. A single wrapper call records every tool, resource, and prompt invocation without touching each handler.
- If you already talk to the model directly and shape the conversation, use Conversations.
- If the model reaches your capabilities through MCP, use the MCP wrapper and get full tool-level coverage for one line.
- If both are true (an app-level agent that also exposes MCP tools), use both. Turns and tool calls link together into one event tree.
Install
The shared shape
Both SDKs expose the same core calls. The naming follows each language’s conventions (snake_case in Python, camelCase in TypeScript), but the behavior is identical.| Call | What it does |
|---|---|
init | Configures the SDK with your API key and optional endpoint. |
begin / end | Opens an interaction for a turn and closes it with the output and success flag. |
track | Records a completed turn in one fire-and-forget call. |
identify | Attaches traits to a user id. |
set_property / set_properties | Adds custom properties to the current interaction (for example model, tokens, cost). |
| MCP wrapper | Auto-captures every tool, resource, and prompt call on an MCP server. |
Full references
Python SDK
Every call with runnable examples.
TypeScript SDK
Every call with runnable examples.
MCP
Wrapping an MCP server in Python and TypeScript.