[Feature Request]: explore ag2 usages in ModelContextProtocol (mcp)

[ohdearquant]

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

Model context protocol provides an interface for easy interaction between clients and server, enabling "client-agnostic" features. For example, custom memory agents, custom tool agents that can be persisted somewhere and be used across any mcp clients, let it be claude desktop or any future applications.

Describe the solution you'd like

integrate fastmcp / fly.io...etc, to allow easy deployment of agents, "agent as a service" (AAAS)

Additional context

• mcp repo: https://github.com/modelcontextprotocol
• fastmcp repo: https://github.com/jlowin/fastmcp
• claude deaktop: https://claude.ai/download

[ohdearquant]

the traditional init chat usage pattern does not work with mcp server due to speed, maybe ag2 should be used as client instead of server

[ohdearquant]

@sternakt @sonichi on the structured output data model from json_schema in #152 , I feel like using mcp servers as validation provider might be a good idea. mcp servers are called with cli, making it easy to put into json schema.

Why?

• dynamic validation,
• separation of concerns of response format and operations on the response

Next steps

• try it to on memory agent, Introduce memory capability #150

[ohdearquant]

in attempt to work out best uses of mcp servers, I created automcp, taking inspiration from fastmcp. A few observations while building this:

1. mcp servers should be quick, providing some kinds of just-in-time service
2. mcp server is incredibly flexible, you can use any callable, just need to define a schema
3. not too sure about the differentiation of prompts/tools/resources, nor its purpose
4. does not support streaming

Issues with existing usage pattern of MCP

mcp now is used to gather context mainly, for example, read into github, read into files. One issue I encountered while using these is that sometimes I do not need the whole file(s) in the LLM conversation context, maybe I just want one piece of information, maybe I just want summary or derivatives. In other words, context management will be useful for improving the use of mcp.

another "issue" or design of MCP is single purposed tools, while this provides modularity, it can increase complexity quickly as more and more service gets connected to the client. Instead of listing out all tools, we can provide a single access point per service, this way service can maintain specific state.

example usage pattern design

{"user": "research on latest papers on XYZ, please implement the algorithms in python"}

{"client": ServiceRequest([
    ExecutionRequest(service=researcher, operation="query_arxiv", content=query1),
    ExecutionRequest(service=researcher, operation="query_arxiv", content=query2),
])}

{"server": "query results..."}
{"client": ServiceRequest([ExecutionRequest(service=researcher, operation="explain", content=...)])}

{"server": "analysis..."}
{"client": ExecutionRequest(service=coder, operation="plan", content=....)}

{"server": "plan..."}
{"client": ServiceRequest([
    ExecutionRequest(service=coder, operation="write_and_debug", content=stage1_part1),
    ExecutionRequest(service=coder, operation="write_and_debug", content=stage1_part2),
    ...
]}

{"server": "codes..., file saved at filepath..."}
{"client": ServiceRequest([
    ExecutionRequest(service=coder, operation="write_and_debug", content=stage2),
    ...
]}
...

{"user": "get these files into ag2 branch XYZ, and create pull request with useful description"}
{"client": ServiceRequest([ExecutionRequest(service=coder, operation="commit_to_gh", content="...")]}

https://github.com/ohdearquant/automcp