Greetings, I’m using LiveKit Cloud Agent with GoHighLevel’s MCP server integration. Despite correct setup and authorization, the agent never calls or uses the MCP and there’s no MCP activity or errors in the traces. Other HTTP tools (e.g., hang-up) work fine. I even added explicit system instructions telling the agent which functions and parameters to call, but it still ignores the MCP. Any ideas how I could fix this? Thanks in advance.

I think you need to call initialize() in entrypoint. When agent starts, in my case, i can see request in my mcp server for tool list request.

Unfortunately i am unable to do such thing since i'm currently using it directly on livekit's cloud, its not self hosted nor am i able to change the code, all i could do is add the mcp and setup the configuration and it should handle it on its own, so i was wondering if there was a trouble with it

My agent is also on cloud. Don't you have agent.py with entrypoint in it which you deploy to cloud?

Correct, but i am unable to modify that code directly, it gets changed dynamically from the configurations i set for the agent in the agent builder tab

Oh. Not sure how that works. But mcp initialisation is required for agent to recognise it. Can explore in this direction.

Right, i noticed that does not happen in the interface, i do not see anything related to the mcp in the code preview at all, even though i added it correctly and configured it as shown in the screenshoot

@crooked-accountant-63197 this looks like an issue with the agent builder to me, I also don't see the code reflecting my specified MCP server. Let me check with the team, but in the mean time you can export your agent as code, https://docs.livekit.io/agents/start/builder/#convert-to-code, and add the MCP server through code (that recipe you linked to is a good place to start for this, https://docs.livekit.io/recipes/http_mcp_client/)

right, i just manually added tools instead for the mean time, can you please let me know when you guys fix it in the agent builder? thank you darryn

Sure, and sorry for the inconvenience. This has apparently been fixed on staging, but not yet deployed to production.