Overview
The Model Context Protocol (MCP) is a standard protocol for communication between LLMs and external services. The MCP connector allows you to connect MCP servers and use their tools as LLM functions.Info: MCP servers provide tools (functions) that can be called by LLMs.
These tools can retrieve data, perform actions, or integrate external
services.
Features Overview
- MCP Server Connection: Connect to MCP servers via HTTP
- Automatic Tool Detection: Tools are automatically retrieved from the server
- Authentication: Support for API Key and OAuth 2.0
- Tool Synchronization: Automatic synchronization of tools from the server
- Test Interface: Integrated tester for testing tool calls
- OAuth Flow: Full support for OAuth 2.0 Authorization Flow
Creating an MCP Configuration
When creating a new MCP configuration, the following options are available:1. Title
Freely selectable name to identify the configuration, e.g., GitHub MCP Server or Custom MCP Service.2. Server URL
The complete URL of the MCP server, e.g.,https://mcp.example.com/api.
Note: The server must support MCP protocol version 2025-03-26 or higher.
3. Transport
Currently only HTTP transport is supported (http).
4. Timeout
Timeout for connections in seconds (default: 30 seconds).5. Authentication
Supported authentication types:-
No Authentication (
none)- For public MCP servers without authentication
-
API Key (
api_key)- API Key: The API key value
- Sent in the
Authorizationheader
-
OAuth 2.0 (
oauth)- Authorization URL: URL for the OAuth authorization endpoint
- Token URL: URL for the OAuth token endpoint
- Client ID: OAuth Client ID (optional, can be dynamically registered)
- Client Secret: OAuth Client Secret (optional)
- Scopes: Desired OAuth scopes
Info: If Authorization URL and Token URL are not provided, an attempt is made to automatically discover them from the server (Authorization Server Discovery).Info: If no Client ID is present and the server supports Dynamic Client Registration, an attempt is automatically made to register a client.
6. OAuth Flow
When using OAuth 2.0:- Start Authorization: Click “Start OAuth” to begin the authorization flow
- Authorization: You will be redirected to the authorization server
- Callback: After authorization, you will be redirected back and the token will be automatically saved
- Token Refresh: Access tokens are automatically renewed if a refresh token is available
Managing Tools
After connecting to an MCP server, the available tools can be synchronized.Tool Synchronization
- Test Connection: Click “Test Connection” to check the connection to the server
- Synchronize Tools: Click “Synchronize Tools” to retrieve all available tools from the server
- Automatic Updates: Tools are automatically updated when the server configuration changes
Tool Details
Each tool contains:- Name: The function name for the LLM
- Description: Description of the function for the LLM
- Input Schema: JSON Schema for the tool parameters
Testing Tools
You can test each tool directly via the integrated test interface:- Select a tool
- Fill in the required parameters
- Click “Test”
- The response is displayed with the result
Providing MCP Tools as Connectors
After creating an MCP configuration and synchronizing tools, you can use these tools as connectors in your assistants.Creating a Connector
- Edit Assistant: Open the assistant configuration
- Add Connector: Add a new connector
- Select MCP Tools: Select the desired MCP tools from the list
- You can select tools from different MCP configurations
- Each tool is provided as a separate LLM function
- Save: Save the connector configuration
Info: MCP tools are automatically converted to LLM functions. The LLM can
call these functions during conversation to retrieve data or perform actions.
Tool Selection
- Multiple Selection: You can select multiple tools from different MCP servers
- Search: Use the search function to quickly find tools
- Filtering: Tools are displayed grouped by MCP configuration
How It Works
When an assistant is configured with MCP tools:- The tools are provided to the LLM as available functions
- The LLM can call these functions based on the user’s request
- The tool calls are forwarded to the corresponding MCP server
- The responses are provided to the LLM and integrated into the conversation
