Add support for MCP servers to Agent Framework, closes #898

docs/agent/configuration.md

@@ -34,18 +34,20 @@ def today() -> str:
     return datetime.today().isoformat()
 
 agent = Agent(
+    model="hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
     tools=[today, wikipedia, arxiv, "websearch"],
-    llm="hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
 )
 ```
 
-## llm
+## model
 
 ```yaml
-llm: string|llm instance
+model: string|llm instance
 ```
 
-LLM path or LLM pipeline instance. See the [LLM pipeline](../../pipeline/text/llm) for more information.
+LLM model path or LLM pipeline instance. The `llm` parameter is also supported for backwards compatibility.
+
+See the [LLM pipeline](../../pipeline/text/llm) for more information.
 
 ## tools
 
@@ -83,7 +85,10 @@ A tool instance can be provided. Additionally, the following strings load tools
 
 | Tool        | Description                                               |
 |:------------|:----------------------------------------------------------|
+| http.*      | HTTP Path to a Model Context Protocol (MCP) server        |
+| python      | Runs a Python action                                      |
 | websearch   | Runs a websearch using the built-in websearch tool        |
+| webview     | Extracts content from a web page                          |
 
 ## method
 

docs/agent/index.md

@@ -41,9 +41,9 @@ def today() -> str:
     return datetime.today().isoformat()
 
 agent = Agent(
+    model="hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
     tools=[today, wikipedia, arxiv, "websearch"],
-    llm="hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-    max_iterations=10,
+    max_steps=10,
 )
 ```
 
@@ -86,31 +86,32 @@ from txtai import Agent, LLM
 llm = LLM("hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4")
 
 websearcher = Agent(
+    model=llm,
     tools=["websearch"],
-    llm=llm,
 )
 
 wikiman = Agent(
+    model=llm,
     tools=[{
         "name": "wikipedia",
         "description": "Searches a Wikipedia database",
         "provider": "huggingface-hub",
         "container": "neuml/txtai-wikipedia"
     }],
-    llm=llm,
 )
 
 researcher = Agent(
+    model=llm,
     tools=[{
         "name": "arxiv",
         "description": "Searches a database of scientific papers",
         "provider": "huggingface-hub",
         "container": "neuml/txtai-arxiv"
     }],
-    llm=llm,
 )
 
 agent = Agent(
+    model=llm,
     tools=[{
         "name": "websearcher",
         "description": "I run web searches, there is no answer a web search can't solve!",
@@ -124,8 +125,7 @@ agent = Agent(
         "description": "I'm a science guy. I search arXiv to get all my answers.",
         "target": researcher
     }],
-    llm=llm,
-    max_iterations=10
+    max_steps=10
 )
 ```
 

setup.py

@@ -27,7 +27,7 @@ extras["dev"] = [
     "pylint",
 ]
 
-extras["agent"] = ["smolagents>=1.13"]
+extras["agent"] = ["mcpadapt>=0.1.0", "smolagents>=1.13"]
 
 extras["ann"] = [
     "annoy>=1.16.3",

src/python/txtai/agent/factory.py

@@ -32,8 +32,8 @@ class ProcessFactory:
             constructor = CodeAgent
 
         # Create model backed by LLM pipeline
-        llm = config.pop("llm")
-        llm = PipelineModel(**llm) if isinstance(llm, dict) else PipelineModel(llm)
+        model = config.pop("model", config.pop("llm", None))
+        model = PipelineModel(**model) if isinstance(model, dict) else PipelineModel(model)
 
         # Create the agent process
-        return constructor(tools=ToolFactory.create(config), model=llm, **config)
+        return constructor(tools=ToolFactory.create(config), model=model, **config)

src/python/txtai/agent/tool/factory.py

@@ -6,7 +6,10 @@ import inspect
 
 from types import FunctionType, MethodType
 
-from smolagents import DuckDuckGoSearchTool, Tool, tool as CreateTool
+import mcpadapt.core
+
+from mcpadapt.smolagents_adapter import SmolAgentsAdapter
+from smolagents import DuckDuckGoSearchTool, PythonInterpreterTool, Tool, tool as CreateTool, VisitWebpageTool
 from transformers.utils import chat_template_utils, TypeHintParsingException
 
 from ...embeddings import Embeddings
@@ -20,7 +23,7 @@ class ToolFactory:
     """
 
     # Default toolkit
-    DEFAULTS = {"websearch": DuckDuckGoSearchTool()}
+    DEFAULTS = {"python": PythonInterpreterTool(), "websearch": DuckDuckGoSearchTool(), "webview": VisitWebpageTool()}
 
     @staticmethod
     def create(config):
@@ -58,8 +61,14 @@ class ToolFactory:
             elif isinstance(tool, str) and tool in ToolFactory.DEFAULTS:
                 tool = ToolFactory.DEFAULTS[tool]
 
+            # Support importing MCP tool collections
+            elif isinstance(tool, str) and tool.startswith("http"):
+                tools.extend(mcpadapt.core.MCPAdapt({"url": tool}, SmolAgentsAdapter()).tools())
+                tool = None
+
             # Add tool
-            tools.append(tool)
+            if tool:
+                tools.append(tool)
 
         return tools
 

test/python/testagent.py

@@ -6,9 +6,11 @@ import os
 import tempfile
 import unittest
 
+from unittest.mock import patch
+
 from datetime import datetime
 
-from smolagents import CodeAgent
+from smolagents import CodeAgent, PythonInterpreterTool
 
 from txtai.agent import Agent
 from txtai.embeddings import Embeddings
@@ -107,3 +109,23 @@ class TestAgent(unittest.TestCase):
 
         self.assertIsNotNone(agent)
         self.assertIsInstance(agent.tools["embeddings1"]("test"), list)
+
+    # pylint: disable=C0115,C0116
+    @patch("mcpadapt.core.MCPAdapt")
+    def testToolsMCP(self, mcp):
+        """
+        Test adding a MCP tool collection
+        """
+
+        class MCPAdapt:
+            def __init__(self, *args):
+                self.args = args
+
+            def tools(self):
+                return [PythonInterpreterTool()]
+
+        # Patch MCP adapter for testing
+        mcp.side_effect = MCPAdapt
+
+        agent = Agent(tools=["http://localhost:8000/mcp"], llm="hf-internal-testing/tiny-random-LlamaForCausalLM", max_steps=1)
+        self.assertEqual(len(agent.tools), 2)