MCP feature injection

[DiscreteTom]

When developing an ACP client, if I want to inject some MCP features (tools, prompts, resources) into the ACP agent, the process is quite complex. I need to:

1. Provide a standalone MCP server.
2. Configure the MCP server for the ACP agent.
3. When spawning ACP agent, I may need to pass additional env vars to make the MCP server communicate with the ACP client process.

As an example, if I build an IDE as an ACP client, and besides fs and terminal actions I want to provide additional tools like getting diagnostics, I have to write a standalone MCP server, configure the MCP server for all configured ACP agents, and when spawning the ACP agents pass an env var like a http server endpoint, the ACP agent will pass the env var to the MCP server, so the MCP server can talk to my main IDE process via the http endpoint.

My propose is that, during the initialization phase, support more client capabilities like:

"clientCapabilities": {
     "tools": [], // i can provide additional tools
      "prompts": [], // i can provide additional prompts
      "resources": [], // i can provide additional resources
    },

Additional benefits: MCP can be configured on the ACP client side, instead of the ACP agent side. This is also useful for remote ACP agent integration if we have local MCP server configured.

Furthermore, we don't even need the fs and terminal client capabilities, just client provided custom tools are enough.