In retrospect, stateful MCP was clearly wrong.
This essentially makes MCP just another REST API endpoint, and lets you use the same infrastructure you already have set up for REST APIs (like load balancers, API gateways, progressive rollouts, etc).
This essentially makes MCP just another REST API endpoint, and lets you use the same infrastructure you already have set up for REST APIs (like load balancers, API gateways, progressive rollouts, etc).
I learnt this with Sun RPC and the whole "The network is the computer".
Somehow this keeps having to be relearnt.
It’s not true. It’s not totally untrue, but don’t hide your head in the sand.
So service API developers want to abstract and hide, over and over, and this means agents have less and less context and do more guessing.
400 Bad Request is fine for a client developer who reads it once at design time and fixes the code forever. it's dead weight for an agent that must self-correct from the string alone. "Expected ISO-8601, got 03/04/2025" is now a functional part of the interface.
Not all agents are developers who can code their own interface.
1. a REST-like API
2. with a spec (OpenAPI, introspection, etc.)
3. with harness-level authz (the ability to allow / deny specific methods in the harness)
The only part that's actually new is #3, but it seems like it should be fairly easy to provide similar support for REST APIs. There might already be a Pi extension that allows you to allow / deny particular method / path patterns for particular sites. Regardless, there's always service-level authz with service accounts, at least for the more sophisticated APIs.
So ya, I don't see why we need a whole new set of standards and protocols just to have harness-level fine-grained API permissions. We could have just done it all with REST.
This is easy to see when you think in terms of a client: It is pretty much impossible to build an opinionated UI for any kind of RESTful API, while creating one for an MCP server is fairly self-explanatory: MCP servers offer a mandatory and complete runtime introspection endpoint (you can retrieve a listing of available tools/resources/prompts etc. along with their parameter and return type schemas). So that means clients have a way to exhaustively describe everything an MCP server is able to do with a vocabulary that carries over exactly to other servers - a tool is a tool everywhere.
There are multiple products that turn OpenAPI schemas into front-ends nowadays.
This is easy to see when you think in terms of a developer. The natural way to adapt an existing REST API into an MCP server while keeping the two synchronized is to simply write a basic translation tool that takes your OpenAPI spec and turns it into an MCP manifest, mapping each HTTP endpoint to an MCP method, and copying over all the request / response type signatures and descriptions for everything. Then, all of the sudden, it hits you: there is no need for this translation layer.
But there's a pretty large body of RFCs that define what makes a good REST API. For instance POST should create things, PUT should re-define them, PATCH should incrementally update them, GET must be side-effect-free, etc. And there's a ton of standard HTTP headers. Pair it with OpenAPI and the whole thing is self-documenting. You can achieve the same effect with gRPC + reflection, or GraphQL + reflection, and there's plenty of mature tooling and interoperability for each system.
By comparison, MCP actually has relatively few conventions. LLMs are very familiar with REST and OpenAPI already. Just use that.
How do you send a big read only query? A POST? A GET hoping nothings logs it or swallows half of it?
How do you deal with delayed answers? Things that most HTTP middleware will timeout long before the result is achieved?
If you have a long running search, the resource is the search. Create a search. Fetch search results. Delete the search when you’re done.
Similarly, if you have a sufficiently complex command, the resource is the command itself. Maybe it requires async work, and you can later fetch the command results.
Anecdotally, this seems to be too big of a leap in mental model for beginners, mostly because all of the REST learning path does is teach you how to use it in CRUD domains.
For delayed answers, MCP offers no improvement, because remote MCP servers use an HTTP transport anyway, and local ones don't involve middleware you can't control. Besides, that's well-trodden ground. SSE, trailers, gRPC streaming, websockets, simple polling, etc. We don't need another protocol.
I maintain https://endara.ai and you can very quickly introduce a ton of capabilities for Claude/Cursor/ChatGPT by configuring all the stuff you use in a single place (and slicing them with profiles if needed).
Most chat products will just have a directory of connectors anyways - so for non technical folks they won't even here the term MCP or REST, they'll just "Install the Gmail plugin".
The problem is often that OpenAPI schemas are often incomplete, since they used to be useful to humans even in an incomplete state.
MCP basically requires you to do the equivalent work of getting your OpenAPI schema complete. It’s just much easier to sell to your managers.
You could use `curl`, or you could bake an HTTP client tool into the agent harness which operates at a more abstract level. The agent would just have a tool that looks a lot like the Fetch API [2] and the harness could simply be configured to inject credentials using the various well-established patterns that have long existed, allow / deny certain methods, etc. The thing is, everything you can do with MCP could also simply be done with REST, so neither is inherently better than the other, I just don't like the proliferation of needless standards.
[1]: https://developer.mozilla.org/en-US/docs/Web/HTTP/Guides/Aut...
[2]: https://developer.mozilla.org/en-US/docs/Web/API/Fetch_API/U...
Besides, the current state of MCP authentication almost always involves just hard coding secrets in .claude/settings.json so it's not like that's doing a great job keeping the creds away from the LLM anyway. All I'm saying is that MCP auth offers no advantage.
It's was simpler to design/implement the AI tool if these interactions had state. When this was generalized to internet interactions this no longer made sense and was gradually cleaned up.
I think at the same time, the AI tool "protocol" to interact with the local system moved away from MCP model.