v0.4 - It is all about the Tools

The core idea behind OpenGPA has always been that the next generation of Agents wasn't so much about 'intelligence' but instead about their access to a wide variety of tools. This is why, from the start OpenGPA had an extensible tool catalog. Unfortunately, until now, adding a tool required development capabilities to create a java extension.

Not anymore! I'm super happy to share this new release, introducing MCP support, a new tool selector, confirmed actions in the UI and a nice improvement for the TTS action.

Built-in support for MCP

A few months back, Anthropic introduced the Model Context Protocol (MCP), a standardized approach to connect agents and tools. This has since then become the standard for Agents, supported by Claude Desktop, OpenAI Operator, and many more. Every single SaaS company is now offering a MCP connector to its service. There are multiple marketplaces with hundreds of connectors available.

And it is now possible to use these tools directly in OpenGPA as well, opening the door to easy native enterprise integration with your favorite tools and services.

{
  "mcpServers": {
    "hubspot": {
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "-e",
        "HUBSPOT_ACCESS_TOKEN=pat-eu1-***",
        "buryhuang/mcp-hubspot:latest"
      ]
    }
  }
}

A new tool selector

In order to facilitate various workflows mixing multiple tools, it is now possible to configure the tools required for a given task. This way, you can ensure your CRM is used instead of a web search, or that information is looked up only in your internal documents, etc...

Once selected, the tools will be available throughout the entire conversation. They can't be changed during the conversation (I might add this later if it's really useful).

Confirmed actions

Because some MCP actions can be critical and require a user review, we have also introduced the possibility to review an action before proceeding. This is available for the email action and will be expanded to other actions on a per-need basis. There is unfortunately no metadata in MCP at the moment to know if an action is just a read or also a write.

Multi-voices Text-To-Speech

This release also packs a few other bugfixes and improvements. In particular, the Text-To-Speech action now supports multiple participants, providing you with NotebookLM capabilities directly into the OpenGPA UI.

Getting started

Running OpenGPA is now as easy as a single command. Adding your custom MCP services still requires some configuration. I have it on the roadmap to find a solution to make this even simpler.

If you want to experiment with self-soveregin AI Agent in your company, do not hesitate to get in touch, I'd be happy to help you out get the most of it and find a balanced approach between ownership and pragmatic resource usage.