Suggestion: MCP-over-ACP

[gkgoat1]

I am making an AI agent system and want to use ACP, but clients should have deep control over the toolset for security (and so we can compile to WASI). This would require a "reverse" MCP transport where, tunneled in ACP traffic, ACP clients act as one or more MCP servers whilst ACP agents act as MCP clients. It would also help with editors like VSCode which have custom MCP APIs and allow ACP security filters and other ACP proxies to fully intercept agent-related traffic.

[benbrandt]

Hey you should check out #243

[jwflicker]

So an IDE can be both an ACP Client can an MCP (HTTP) Server. It can forward itself as an MCP Server to the ACP Agent when creating a new session. The combination IDE/ACP Client/MCP Server can dynamically add/remove tools and resources to/from its MCP Server while the ACP Agent is connected it, and the ACP Agent will get notified and adapt. The MCP Server can also make its own side-prompts using MCP Sampling. I haven't looked into sampling... i wonder if it can be invoked at any time, independent of a Tool call, to, for example, to provide Content Assist...

So we are using this in an IDE that runs an HTTP MCP Server in a thread of its own process. This lets the agent call tools based on chat context to get compilation errors, remote files, interact with middleware connections configured with the IDE using IDE-secured credentials such as making SQL calls to databases, interact with the UX, and other stuff using tool calls. It does (probably) require a dedicated HTTP port for the MCP Server to run on.

There are some limitations we raise as we incrementally pursue this approach such as:
MCP Resource, Templates, Elicitation thru ACP #342