Feature Request: MCP Server Integration in Langfuse Playground

[rahulreddy1109-lang]

Describe your question

Feature Request

I'd like to request support for integrating MCP (Model Context Protocol) servers directly into the Langfuse Playground.

Current State

Currently, the Playground supports:

• Tool calling with JSON schema definitions
• Mocking tool responses for testing

The Langfuse MCP server exists for exposing prompts to MCP clients (Claude Desktop, Cursor), but there's no way to connect custom MCP servers as tools within the Playground itself.

Desired Functionality

It would be valuable to:

• Connect custom MCP servers to the Playground
• Use MCP tools alongside the existing JSON schema tool definitions
• Test prompts that rely on MCP tools in real-time
• Configure Azure OpenAI (or other providers) with MCP tool support

Use Case

I want to test prompts in the Playground that use custom MCP tools, similar to how I can currently test with JSON schema-defined tools, but leveraging the full MCP protocol for more complex tool integrations.

Langfuse Cloud or Self-Hosted?

Self-Hosted

If Self-Hosted

3.162.0

If Langfuse Cloud

No response

SDK and integration versions

No response

Pre-Submission Checklist

• I have checked for existing issues/discussions and consulted Langfuse AI.

[inkeep-ai-assistant[bot]]

Currently, the Langfuse Playground supports tool calling with JSON schema definitions and allows you to test prompts by mocking tool responses—but it does not support connecting actual MCP servers as executable tools. (1) (2)

The existing Langfuse MCP server is designed for the opposite direction: it exposes Langfuse-managed prompts to MCP-compatible clients like Claude Desktop or Cursor, rather than allowing the Playground to consume external MCP tools. (3)

This feature request aligns with an existing discussion. As noted by a Langfuse team member, tools added to the Playground currently "can only simulate the tool call activity" and "it is not possible to add an actual MCP server here yet." However, this capability is on the roadmap. (4)

Workaround options:

If you need to integrate custom tool logic now, you could:

1. Expose your MCP server as an OpenAI-compatible endpoint (using something like LiteLLM as a proxy), then configure Langfuse to point at that endpoint. If your server isn't publicly accessible, you can use a tunnel like ngrok. (4)

2. Build a backend bridge that receives tool calls from the Playground (defined via JSON schema), makes HTTP requests to your MCP server, and transforms responses to match Langfuse's expected output format. This requires custom development. (4)

For Azure OpenAI or other providers, you can configure LLM connections in Project Settings > LLM Connections, and the Playground supports any provider implementing the OpenAI API schema. (5)

I'd recommend adding your use case to the existing GitHub discussion to help prioritize this feature: MCP server as tool in playground

📚 Sources:

1. LLM Playground
2. Tool Calling and Structured Output in Playground - Langfuse
3. Langfuse MCP Server
4. MCP server as tool in playground
5. LLM Connections

---

Have another question? Just tag @inkeep.

[jannikmaierhoefer]

Thanks @rahulreddy1109-lang! We are tracking this here: https://github.com/orgs/langfuse/discussions/6451