traceAI-mcp-use

# traceAI-mcp-use OpenTelemetry tracing for [mcp-use](https://pypi.org/project/mcp-use/) agents (`mcp_use.MCPAgent`), sent to Future AGI. `FutureAGICallback` is a LangChain callback handler. You pass it to the agent with `MCPAgent(..., callbacks=[...])`. It records one span for each agent run, each LLM call and each tool call. It patches nothing, imports neither `langfuse` nor `traceai_mcp`, and sets no environment variables. This package traces the agent loop. [`traceAI-mcp`](../mcp) (`traceai_mcp`) works at the MCP transport level and is a different layer: a green transport test does not show the agent run, its model calls or its tool calls. See [Other tracing in the same process](#other-tracing-in-the-same-process). ## Installation ```bash pip install traceAI-mcp-use ``` It accepts `mcp-use>=1.7.1,<2`, `langchain-core>=1.0.0,<2` and `fi-instrumentation-otel>=1.1.0`, on Python `>=3.11,<3.14` (mcp-use itself requires 3.11). The suite runs with mcp-use 1.7.1 (the latest release on 2026-10-16) on Python 3.11, 3.12 and 3.13, and was run end to end against a real model and a real stdio MCP server with mcp-use 1.7.1, langchain-core 1.6.12, langchain-openai 1.7.2 and langgraph 1.2.16. ## Usage ```python from fi_instrumentation import register from fi_instrumentation.fi_types import ProjectType from mcp_use import MCPAgent, MCPClient from traceai_mcp_use import FutureAGICallback trace_provider = register(project_type=ProjectType.OBSERVE, project_name="mcp-use-agent") client = MCPClient.from_dict({"mcpServers": {"calculator": {"command": "python", "args": ["server.py"]}}}) agent = MCPAgent( llm=llm, # any LangChain chat model client=client, callbacks=[FutureAGICallback(tracer_provider=trace_provider)], ) result = await agent.run("What is 2 + 3?") ``` Pass the provider `register()` returns: `register()` does not set the global provider unless you ask it to, and without `tracer_provider` the callback uses the global one. One callback can serve many a…