Tool Agent README

======================

Overview

Tool Agent is a Python package that converts any command-line tool into a Large Language Model (LLM) agent. This allows you to interact with the tool using natural language, making it easier to use and integrate with other applications.

Requirements

• Python 3.12 or later
• FastAPI
• Requests
• Pydantic
• Uvicorn
• cwltool
• mcp (Model Context Protocol SDK)

Installation

To install Tool Agent, run the following command:

pip install git+https://github.com/hubentu/cmdagent

How the Framework Works

Tool Agent leverages the Model Context Protocol (MCP) to bridge command-line tools and Large Language Models (LLMs). The framework works by converting CWL (Common Workflow Language) tool definitions into MCP-compatible agents that can be discovered and invoked by LLMs through natural language queries. Here's how it works: you create an MCP server instance using mcp_api, register your domain-specific tools by providing their CWL definitions via add_tool(), and then start the server. The MCP server exposes these tools as discoverable agents that any MCP-compatible client (like Cursor) can query and invoke. When an LLM needs to use a tool, it queries the MCP server for available tools, selects the appropriate one, and invokes it with the necessary parameters. The tool executes within a containerized environment (as specified in the CWL), processes the request, and returns results back through the MCP protocol to the LLM, which then presents the answer to the user in natural language.

Usage

MCP server

The framework allows you to set up an MCP server with predefined tools for specific domains. For example, to create a bioinformatics-focused MCP server, you can use the following setup (as shown in examples/bioinfo_question.py):

from cmdagent.mcp_api import mcp_api

mcp = mcp_api(host='0.0.0.0', port=8000)
mcp.add_tool('examples/ncbi_datasets_gene.cwl', 'ncbi_datasets_gene')
mcp.add_tool('examples/bcftools_view.cwl', 'bcftools_view', read_outs=False)
mcp.serve()

This creates an MCP server that exposes two bioinformatics tools:

• ncbi_datasets_gene: Retrieves gene metadata from NCBI datasets
• bcftools_view: Subsets and filters VCF/BCF files

Once the server is running, you can configure your MCP client (e.g., in Cursor) to connect to it:

{
    "mcpServers": {
        "cmdagent": {
            "url": "http://localhost:8000/mcp",
            "transport": "streamable-http"
        }
    }
}

With this setup, you can ask the LLM natural language questions like:

• "Give me a summary about gene BRCA1"
• "Subset variants in the gene BRCA1 from the https://ftp.ncbi.nlm.nih.gov/pub/clinvar/vcf_GRCh38/clinvar.vcf.gz"

The LLM will automatically discover the available tools, understand their parameters, invoke the appropriate tool with the correct arguments, and present the results in a user-friendly format.

• Start MCP server

python examples/bioinfo_question.py

• Call by MCP client from Cursor

Function call

• Creating an API

To create an API, import the tool_api function from cmdagent.remote_api and pass in the path to a CWL file and the name of the tool:

from cmdagent.remote_api import tool_api

api = tool_api(cwl_file='tests/dockstore-tool-md5sum.cwl', tool_name='md5sum')
api.serve()

The api.serve() method will start a RESTful API as a service, allowing you to run the tool remotely from the cloud or locally.

• Creating a Tool Agent

To create a tool agent, import the cmdagent function from cmdagent.agent and pass in the API instance:

from cmdagent.agent import tool_agent

ta = tool_agent(api)
md5 = ta.create_tool()
md5(input_file="tests/dockstore-tool-md5sum.cwl")

Function md5 is created automatically based on the api.

• Function call with Gemini

To integrate the tool agent with Gemini, import the GenerativeModel class from google.generativeai and create a new instance:

import google.generativeai as genai

genai.configure(api_key="******")
model = genai.GenerativeModel(model_name='gemini-1.5-flash', tools=[md5])

chat = model.start_chat(enable_automatic_function_calling=True)
response = chat.send_message("what is md5 of tests/dockstore-tool-md5sum.cwl?")
response.text

'The md5sum of tests/dockstore-tool-md5sum.cwl is ad59d9e9ed6344f5c20ee7e0143c6c12. \n'