Hey. I got it running with a selfhosted MCP, but I wanna use SSE, since it can push messages to LibreChat or any client (is that actually the case??). I connected LibreChat to the remote server, I can see tools, and a LLM is able to call them (with valid response from server). The problem is, that it is not able to see which parameter it should send, or isnt aware of any structure of the request. For example, when it calls create_call function, it just makes stuff up about the rest. I believe it only knows the name of the tools. What could be the problem?

It sounds like LibreChat is not injecting the full tool schema (including parameter definitions) into the LLM's context—only the tool names. This usually happens if the client does not parse or forward the tool definitions from the MCP server response. Actionable step: Check that your client is passing the entire tool schema to the LLM. For example, the tool definition for

create_call

should look like:

json
{
  "type": "function",
  "function": {
    "name": "create_call",
    "parameters": {
      "type": "object",
      "properties": {
        "assistantId": { "type": "string" },
        "phoneNumberId": { "type": "string" },
        "customer": { "type": "object", "properties": { "phoneNumber": { "type": "string" } } }
      }
    },
    "description": "Creates an outbound call"
  }
}

Verify that this structure is present in the payload sent to your LLM. If not, update your integration to include the full tool schema from the MCP server response, not just the tool names. [See MCP integration details](https://docs.vapi.ai/tools/mcp#how-mcp-works) Sources: - [Model Context Protocol (MCP) Integration – How MCP Works](https://docs.vapi.ai/tools/mcp#how-mcp-works) - [Vapi MCP Server – Core Tools and Example Payloads](https://docs.vapi.ai/sdk/mcp-server#core-tools)

Hey, Can you send me the latest call_id? Along with that, can you provide me the duration in the recording where you had this issue?

call ID: 07767ca0-69e3-42dd-86ac-5cdc38da6a1d

Yes, you should be able to utilize the MCP tool to collect the variable data at the very beginning of the call so the assistant can access it. You can read more information about dynamic variables here: