New Feature: Tools Vs Resources

[ouarrho]

Is your feature request related to a problem? Please describe.

Voltagent is limited because it focuses solely on "tools" for actions. It lacks "resources," which are crucial for the AI to access and utilize information, thus hindering its ability to understand context and provide comprehensive responses.

Describe alternatives you've considered

No response

Additional context

No response

Describe the thing to improve

It would be highly beneficial to enhance the AI framework by integrating both "tools" and "resources." While "tools" enable the AI to perform actions, "resources" would provide the AI with access to information.

[omeraplak]

Great point, and thanks again for bringing this up!

Do you happen to know any existing frameworks that already support a clear separation between “tools” and “resources”? We’d love to take a closer look at how they structure this and explore how a similar concept could be integrated into VoltAgent.

It’s definitely an interesting direction, and we’re open to contributions if you’d like to collaborate or prototype something together!

[ouarrho]

Thank you again @omeraplak for your quick reply,
I see the resource concept on the Model Context Protocol (MCP) website.

While other frameworks like LangGraph and Kaiban handle similar concepts (though not using the exact terms "tools" and "resources").

And ofcourse I'd like to contribute to VoltAgent it seems like a great project. Let me know how to help.

[omeraplak]

That’s super helpful! Thanks again for the insights and the MCP reference!

To help us move forward, could you share your thoughts on a few things?
• What kind of resources do you think agents might need?
• How would you explain the difference between a tool and a resource in your mind?
• Do you have an example or two where a resource would be better than a tool or retriever?
• How do you imagine using a resource inside a VoltAgent agent?

No pressure to write anything big, even quick answers here are more than enough. Happy to explore this together!

[ouarrho]

The concept of a resource is better understood in the context of docs/db-reading/file-reading/JSON, etc.

Otherwise, a tool can perform actions like inserting data, calling APIs, making calculations, etc.

Therefore, our agents can have many types: tool, resource, etc.

Here's an example to illustrate the benefit of this concept: imagine an agent that needs to read the API documentation of a backend (Swagger). The idea is simple: add the Swagger JSON as a resource (a JSON file in this case), and tadaaaa 🎉, our agent now has knowledge about our backend documentation.

Another example: let's assume we have a list of actions (fetch, click, navigate, mousemove, etc.) and we need our agent to have knowledge about these actions. It can simply access them via a resource. Overall, a resource is a source of data that the agent can access; it can be a JSON, text, XML, or whatever.

I see really big potential in this project and I'm open to a discussion or even a call session to discuss various topics related to AI and its impact on the near future and how we can improve this project to the next level, so feel free to contact me you can find my infos on my profile.

Cheers🎉

[xmile1]

Hi, thanks for this awesome library!

Is full MCP protocol support (resources and prompts) still on the radar? By now, the value of these additional MCP entities has become very evident, and I was wondering if supporting them with their full specification is something you're considering, or if this is already possible and I just haven't figured it out.

Right now, to use an MCP server, I retrieve the tools available from the server and pass them to my agent. However, this approach excludes resources and prompts that may be available on the server.

Current behavior:

const mcpConfig = new MCPConfiguration({ servers: {...} });
const tools = await mcpConfig.getTools(); // Only gets tools
const agent = new Agent({ tools }); // Resources and prompts are not accessible

Expected behavior:
Access to all three MCP capabilities:

• ✅ Tools (currently working)
• ❓ Resources
• ❓ Prompts

Would love to know if there are plans to expose these capabilities, or if there's an existing pattern I'm missing!