MCP server auto search and add to toolkit

examples/conversation_with_react_agent/code/react_agent_with_auto_mcp_toolkit.py

@@ -0,0 +1,88 @@
+# -*- coding: utf-8 -*-
+"""An example of how to use AutoServiceToolkit with ReAct agent."""
+import os
+from agentscope.agents import ReActAgent, UserAgent
+import agentscope
+from agentscope.service import ServiceToolkit
+
+if __name__ == "__main__":
+    YOUR_MODEL_CONFIGURATION_NAME = "dashscope_config"
+    YOUR_MODEL_CONFIGURATION = {
+        "model_type": "dashscope_chat",
+        "config_name": YOUR_MODEL_CONFIGURATION_NAME,
+        "api_key": os.getenv("DASHSCOPE_API_KEY"),
+        "model_name": "qwen-max",
+    }
+
+    agentscope.init(
+        model_configs=YOUR_MODEL_CONFIGURATION,
+        project="Conversation with ReActAgent",
+        save_api_invoke=True,
+    )
+
+    # Prepare the toolkit that can automatically search tools for the agent
+    service_toolkit = ServiceToolkit(
+        mcp_search_allow=True,
+        # If you want the agent control the MCP server installation, uncomment
+        # auto_install_mcp=True
+    )
+
+    human_confirm_sys_prompt = (
+        "You are a helpful assistant that can help people finish different "
+        "tasks."
+        "When you need new tools, you can automatically search for appropriate"
+        " tools (in terms of MCP servers)."
+        "You need to follow the the following instructions: \n"
+        "1. When you obtain some candidate MCP servers, you need to "
+        "analysis the candidate MCP servers, then break the reasoning process "
+        "by calling the `finish` function and wait for human response.\n"
+        "2. When the user confirms to install a MCP server, you need "
+        "to obtain the configuration template of the MCP server. \n"
+        "3. When you get the template, you can break the reasoning process by "
+        "calling the `finish` function and "
+        "ask the user to confirm the configuration for you. \n"
+        "4. With the confirmed MCP configuration, you can go ahead and "
+        "install the MCP server.\n"
+        "5. After all above, you can use the tools to solve user's original "
+        "question."
+    )
+
+    # when auto_install_mcp=True
+    auto_confirm_sys_prompt = (
+        "You are a helpful assistant that can help people finish different "
+        "tasks."
+        "When you need new tools, you can automatically search for appropriate"
+        " tools (in terms of MCP servers)."
+        "You need to follow the the following instructions: \n"
+        "1. When you obtain some candidate MCP servers, you need to analysis "
+        "the candidate MCP servers, and select the best one to install. \n"
+        "2. You may need to get the configuration template of the MCP server "
+        "before install it.\n"
+        "3. After all above, you can use the tools to solve user's original "
+        "question."
+        "4. If you encounter errors when using tools (e.g., invalid API key) "
+        "you can ask the user to provide necessary information."
+    )
+
+    # Example query 1: "I want to get today's latest new about US tariff"
+    # Example query 2: "help me generate a fasting plan with your knowledge in
+    # a markdown file, write it to the directory /Users/xxxx/Desktop"
+
+    agent = ReActAgent(
+        name="assistant",
+        sys_prompt=human_confirm_sys_prompt,
+        # when auto_install_mcp=True, use the auto_confirm_sys_prompt
+        # sys_prompt=auto_confirm_sys_prompt,
+        model_config_name=YOUR_MODEL_CONFIGURATION_NAME,
+        verbose=True,
+        service_toolkit=service_toolkit,
+    )
+    user = UserAgent(name="User", input_hint="User Input ('exit' to quit): ")
+
+    # Build
+    x = None
+    while True:
+        x = user(x)
+        if x.content == "exit":
+            break
+        x = agent(x)