# Langfuse Prompt Management MCP Server

[Model Context Protocol](https://github.com/modelcontextprotocol) (MCP) Server for [Langfuse Prompt Management](https://langfuse.com/docs/prompts/get-started). This server allows you to access and manage your Langfuse prompts through the Model Context Protocol.


## Features

### MCP Prompt

This server implements the [MCP Prompts specification](https://modelcontextprotocol.io/docs/concepts/prompts) for prompt discovery and retrieval.

- `prompts/list`: List all available prompts

  - Optional cursor-based pagination
  - Returns prompt names and their required arguments, limitation: all arguments are assumed to be optional and do not include descriptions as variables do not have specification in Langfuse
  - Includes next cursor for pagination if there's more than 1 page of prompts

- `prompts/get`: Get a specific prompt

  - Transforms Langfuse prompts (text and chat) into MCP prompt objects
  - Compiles prompt with provided variables

### Tools

To increase compatibility with other MCP clients that do not support the prompt capability, the server also exports tools that replicate the functionality of the MCP Prompts.

- `get-prompts`: List available prompts

  - Optional `cursor` parameter for pagination
  - Returns a list of prompts with their arguments

- `get-prompt`: Retrieve and compile a specific prompt

  - Required `name` parameter: Name of the prompt to retrieve
  - Optional `arguments` parameter: JSON object with prompt variables

- `update-prompt`: Create new prompt versions or update existing ones, with production promotion
  - Required `name` parameter: Name of the prompt to create or update
  - Optional `prompt` parameter: The prompt content (string for text prompts, array of message objects for chat prompts). Required when creating new versions, omit when only updating labels.
  - Optional `version` parameter: Specific version number to update labels on. If provided, only labels will be updated. If omitted, creates a new version.
  - Optional `type` parameter: Type of prompt ("text" or "chat", defaults to "text")
  - Optional `config` parameter: Configuration object for model parameters (e.g., temperature, model name)
  - Optional `labels` parameter: Array of labels to assign to this prompt version (e.g., ["staging", "production"])
  - Optional `tags` parameter: Array of tags to categorize the prompt
  - Optional `promoteToProduction` parameter: Boolean to automatically add "production" label (defaults to false)

## Development

```bash
npm install

# build current file
npm run build

# test in mcp inspector
npx @modelcontextprotocol/inspector node ./build/index.js
```

## Usage

### Step 1: Build

```bash
npm install
npm run build
```

### Step 2: Add the server to your MCP servers:

#### Claude Desktop

Configure Claude for Desktop by editing `claude_desktop_config.json`

```json
{
  "mcpServers": {
    "langfuse": {
      "command": "node",
      "args": ["<absolute-path>/build/index.js"],
      "env": {
        "LANGFUSE_PUBLIC_KEY": "your-public-key",
        "LANGFUSE_SECRET_KEY": "your-secret-key",
        "LANGFUSE_BASEURL": "https://cloud.langfuse.com"
      }
    }
  }
}
```

Make sure to replace the environment variables with your actual Langfuse API keys. The server will now be available to use in Claude Desktop.

#### Cursor

Add new server to Cursor:

- Name: `Langfuse Prompts`
- Type: `command`
- Command:
  ```bash
  LANGFUSE_PUBLIC_KEY="your-public-key" LANGFUSE_SECRET_KEY="your-secret-key" LANGFUSE_BASEURL="https://cloud.langfuse.com" node absolute-path/build/index.js
  ```

## Limitations

The MCP Server is a work in progress and has some limitations:

- Only prompts with a `production` label in Langfuse are returned
- All arguments are assumed to be optional and do not include descriptions as variables do not have specification in Langfuse
- List operations require fetching each prompt individually in the background to extract the arguments, this works but is not efficient