I’ve been working with MCP for a bit and really like the model it encourages: explicit tools, clear schemas, tight boundaries. It works well when you’re running your own agent or backend that bridges the model to an MCP server.
One thing I keep running into, though, is that this doesn’t translate well to chat-style workflows.
To be clear upfront: I’m not asking for the model to have general network access, hit arbitrary URLs, or do anything autonomous in the background.
What I’m talking about is something much more constrained — closer to function calling — but backed by MCP-style tools.
Right now, if I want to experiment or do anything interactive with real systems, I have to stand up a local or remote proxy that:
-
talks to OpenAI,
-
then talks to an MCP server,
-
and then feeds results back into the conversation.
That works, but it adds a lot of friction, especially for exploratory or human-in-the-loop workflows.
It feels like MCP already solves most of the hard parts here:
-
tools are explicit
-
schemas are known
-
calls are structured
-
boundaries are clear
So I’m curious whether anyone else has thought about, or run into, the same gap:
chat is great for reasoning and iteration, MCP is great for real capabilities, but the two don’t quite meet unless you build your own bridge.
A user-approved, scoped way for chat to invoke predefined MCP tools (no raw sockets, no dynamic endpoints, no background execution) feels like it could unlock a lot of interesting use cases without blowing up the security model.
Examples that come to mind:
-
checking the state of a system
-
triggering a narrowly scoped action
-
querying internal APIs or services
-
experimenting with agent behavior before hardening it for production
Maybe this already exists and I’m missing it, or maybe everyone is just rolling their own proxies today.
Curious how others are handling this, and whether something more native is even on the roadmap.