Local Agentic Workflow MCP Server

This MCP server allows you to run the Agentic Workflows using a local LLM server using Ollama CLI. This has been tested for VS Code. The workflows are designed as follows:

Parallel Agentic Workflow

1. Orchestrator Agent: This agent is the first to be called. It is responsible for orchestrating the workflow and calling the other agents as needed, depending on the user prompt.
2. Worker Agents: The orchestrator agent calls these agents to perform specific tasks. The Orchestrator Agent calls these agents in parallel and waits for their responses before proceeding to the next step.
3. Aggregator Agent: This agent is called last to aggregate the results from the other agents and return the final result to Github Copilot.

Sequential Agentic Workflow

1. All the agents are called sequentially, and the output of one agent is passed as input to the next agent. The final result is returned to Github Copilot. These agents are called in the order they are defined in the config file.

Prerequisites

- Python 3.8 or higher
- pip (for installing Python packages)
- Ollama CLI (for local LLMs)

Installation

1. Clone the repository:

   git clone https://github.com/Antarpreet/agentic-workflow-mcp.git
   

2. Start an LLM server using the ollama CLI. For example, to start the deepseek-r1:14b model, run:

   ollama run deepseek-r1:14b
   

   This will start the LLM server on http://localhost:11434 by default.

3. Install the required Python packages:

    pip install -r requirements.txt
   

4. Add MCP Server to VS Code:

   • Open the Command Palette (Ctrl+Shift+P or Cmd+Shift+P on macOS).
   • Type MCP and select MCP: Add MCP Server.
   • Select Command (stdio) as the server type.
   • Enter the command to start the MCP server.

   python -m uv run mcp run agentic-workflow-mcp/server.py
   

   • Name the server: Agentic Workflow.
   • Select User settings to add the server to all workspaces or Workspace settings to add it to the current workspace only.
   • If you already have other workspaces open, restart VS Code for changes to take effect.
   • This will open the settings.json file with the new MCP server configuration, which should look like this:

   "mcp": {
       "servers": {
           "Agentic Workflow": {
               "type": "stdio",
               "command": "python",
               "args": [
                   "-m",
                   "uv",
                   "run",
                   "mcp",
                   "run",
                   "agentic-workflow-mcp/server.py"
               ]
           }
       }
   }
   

5. Add Config for the MCP Server as follows:

   • Adjust the default config in config.json as needed. The config settings are detailed further below.

   • Copy the server folder to the user folder. (C:\Users\<username>\agentic-workflow-mcp on Windows or ~/agentic-workflow-mcp on Linux/MacOS).

     Windows:

     xcopy /E /I agentic-workflow-mcp %homedrive%%homepath%\agentic-workflow-mcp
     

     Mac/Linux:

     rm -rf ~/agentic-workflow-mcp
     cp -r agentic-workflow-mcp ~/agentic-workflow-mcp
     

   • Anytime you make any changes to these files, copy them to the user folder again and restart the MCP server in the settings.json file for the changes to take effect.

6. Start the MCP server:

   • Click the start button above the MCP server configuration in the settings.json file.
   • This will start the MCP server; you can see the logs in the Output panel under MCP: Agentic Workflow by clicking either the Running or Error button above the MCP server configuration.

7. Start using the MCP server:

   • Open GitHub Copilot in VS Code and switch to Agent mode.
   • You should see the Agentic Workflow MCP server and start_workflow tool in the Copilot tools panel.
   • You can now use the start_workflow tool to start the Agentic Workflow. Prompt example:

   Use MCP Tools to start a workflow to YOUR_PROMPT_HERE.
   

Config Settings

The config settings are defined as follows

Agents are defined as follows

Custom Embeddings

Local LLM Tools

GitHub Copilot Tools

start_workflow

This tool is used to start the Agentic Workflow. It takes a prompt as input and returns the result of the workflow.

Example usage (Parallel Agentic Workflow)

The example configuration contains agents that get information regarding a country. There are three agents besides the Orchestrator Agent and Aggregator Agent: Country, Flag, and Language. Depending on your prompt, each of them will return their responses, and the combined responses are sent to the Aggregator Agent, which will return the final result.

There is also a prompt.txt file, which contains the prompt for the Orchestrator Agent as an example to show how to use a file instead of a string for the prompt. The prompt.txt file is used in the prompt_file key of the Orchestrator Agent in config.json.

Test Prompt: Use MCP tools to start workflow to confirm whether France is real or not

Troubleshooting

• If the MCP server is not making any requests to the LLM server, do the following:

  1. Restart VS Code as a sanity check.
  2. Sometimes the Ollama server doesn't respond if it hasn't been used for a while. In that case, open the ollama URL (http://localhost:11434) in a browser. You should see the message Ollama is running. If you don't see this message, restart the Ollama server using the command ollama run deepseek-r1:14b.
  3. Copy the server files again using the commands above, another sanity check.
  4. If your request to the local LLM is failing or taking too long, it might be a good idea to increase the read timeout in the server.py file.
  5. Restart the MCP server in the settings.json file.
  6. Create a new chat in GitHub Copilot.

• To see the internal logs, you can set the verbose key in the config.json file to true. This will include all the logs in the response. You can see these by expanding the start_workflow tool in the Copilot chat window.

• You can also check the logs using:

  Windows:

  type %homedrive%%homepath%\agentic-workflow-mcp\logs.txt
  

  Mac/Linux:

  tail -f ~/agentic-workflow-mcp/logs.txt
  

TODO

• Add support for multi-modality using local LLM servers.