Building MCP servers for ChatGPT and API integrations
platform.openai.com69 points by kevinslin 2 days ago
69 points by kevinslin 2 days ago
I find the state of MCPs with OpenAI quite confusing:
- As an end-user, you can connect MCPs only with `search` and `fetch` that only work in deep research mode.
- As a developer, you can use MCP with the API and that supports the full set of MCP tools - all tools become available; this shows in the dev playground.
- For Custom GPTs, they support any action, but not MCPs. So if you had a layer to translate MCP to their API spec, it will work. But Custom GPTs with actions only support models 4o and 4.1; so you don't get the benefit of the o-series of models.
Figuring out what works when is harder than it needs to be.
OpenAI has fallen behind, way behind. Stop trying to benchmark off openai, it's pointless imo. They're not even playing the same game as everyone else.
The lack of MCP support in their desktop client is especially disappointing considering everything Anthropic has shipped on Claude Code the last few weeks.
ChatGPT desktop client with only search/fetch MCPs is far, far inferior to CC from a utility/value perspective.
It feels like they don’t have a good product strategy that matches their engineering throughput.
Has something changed? That seems to be the same page they've had for a couple of months. It's only for deep research mode, and is restricted to Pro and Enterprise.
I thought the same, a possible change from past might be the detailed data leakage and attack explanations?
Eg how I described here a while ago: https://x.com/wunderwuzzi23/status/1930899939737166075?s=46&...
Ironically, I have a blog post drafted that explains this also in detail, and should probably still publish it.
Same. I thought it’s now available for Plus, but apparently not the case
It’s disappointing they are gating this and browser agent for $100 users, and even for 100, it’s only two tool methods.
The support here is really weird.
If I understand correctly, it requires your MCP server to have exactly two tools - search and fetch.
So this is not really support for MCP in general, as in all the available MCP servers. It’s support for their own custom higher-level protocol built on top of MCP.
For ChatGPT and DeepResearch yes, not when using the API. I guess you could just return empty results if you want to offer other tools as well (can't test it now, since custom connectors only supports Workspace or PRO accounts for this moment).
Quote we're talking about: > To work with ChatGPT Connectors or deep research (in ChatGPT or via API), your MCP server must implement two tools - search and fetch.
Reference links:
- Using remote MCP servers with the API: https://platform.openai.com/docs/guides/tools-remote-mcp
- Which account types can setup custom connectors in ChatGPT: https://help.openai.com/en/articles/11487775-connectors-in-c...
this guide is just for an example of how to build a single mcp (e.g. a vector store). chatgpt connectors implement mcps in general now
I don’t believe this is the case. Do you have a link for that?
Cause from TFA “To work with ChatGPT Connectors or deep research (in ChatGPT or via API), your MCP server must implement two tools - search and fetch.”
Also, this page is actually the only docs site about MCP they have, and their help articles link to it too.
It is very nice to get MCP support in ChatGPT. OpenAI really fumbled the bag with the OpenAPI-based Custom Actions (or was it Custom GPTs?). The web editor experience was always incredibly buggy, even months after initial release. MCP servers allow us to move nearly all of the tool definition bits into the server codebase itself, so we can change things on the fly / version control / feature flag tools etc. much better.
Is it safe to say that MCP has "won" vs. A2A? Or is this a misreading of the situation?
They're not directly solving the same problem. MCP is for exposing tools, such as reading files. a2a is for agents to talk to other agents to collaborate.
MCP servers can expose tools that are agents, but don't have to, and usually don't.
That being said, I can't say I've come across an actual implementation of a2a outside of press releases...
I think they have different use cases. MCP is for tool calling, A2A for agents communicating between themselves.
I like how they plugged the MCP support into their existing tools. Pretty smart!
I love how they don't actually explain why I (or anyone else) would ever implement their API.
Given how disastrous the AI 'industry' has been, between misappropriating data from customers, performing actions on behalf of customers that lead to data and/or financial loss, and then seeking protection from the law in one or more cases of these, isn't providing an MCP service essentially requiring you to notify customers of a GDPR-or-similar data compromise event at some point in the future when it suddenly but inevitably betrays you?
Like, isn't OpenAI just leading people to a footgun and then kindly asking them to use it, for the betterment of OpenAI's bottom line, which was significantly in the red for FY24?
Ah yes, “disastrous” not quoted but felt the need to put into quotation “industry” as if it’s not the biggest shift that has happened in your entire lifetime. That is intellectually dishonest and incongruent with the reality that you live in
They said the same thing about crypto. We're still transacting in fiat.
i guess the big difference is normal people are using LLM’s now in the form of google search AI overview, chatgpt, etc. Do you know “normal” people that use crypto for buying and selling goods?
I know 0. And i work in tech lol.
That's not the "big difference" you're looking for, though. I'm the tech person in the lives of many normal people, and they ask me, "Hey, why does Google search suck now? How do I turn off this AI thing that keeps giving wrong information?", as if I know some magic secret switch to turn that off (adding udm=14 to the query URL is too difficult for normal people).
I don't hear anything from people I actually know who are exposed to this stuff being like, "Oh my God, this has massively improved my life and I can't live without it!" No its stuff like, "How do I uninstall Siri off my iPhone? It doesn't work right anymore and can't understand me, and keeps turning itself back on when there is an iOS update."
I don't think any amount of marketing is going to change the fact that the service doesn't work and isn't up to user expectations.
It's undercooked and now its too late to put it back in the oven, so to speak.
Google search has sucked for years, even before the AI stuff was added. To be brutally honest. My personal complaint with google search is that the AI overview is pushed prominently. The overview itself is not bad, especially if you’re just trying to look up something quickly that’s not that important. For example “which two colors make purple?” or “how far is NY from SF?” Expecting it to be 100% factually correct is an issue, and should not be relied on. But google doesn’t prominently point out that AI can make mistakes. Hence the user expectation problem…
And siri has been junk for years too. Not sure what LLM’s have to do with that. Siri activates via a machine learning model that runs on device, which listens for a “wake word”, and that tech has been around for 10 years at this point. It’s possible apple messed up their machine learning model in a recent update? Idk :) But that’s unrelated to LLM’s.
I agree hype has gone out of control! the tech companies are over-promising. Expectations are too high, like you said. That doesn’t mean LLM’s are useless like crypto though. It’s early days still. Crypto is not useful to _any_ normal person, unless you live in a handful of countries with crazy inflation, or you’re a crypto nerd, or you’re trying to move dirty money around.
But all of this is useful to almost anyone:
- quickly summarize a recipe online, removing all the junk ads from the website.
- Brainstorming a quick travel itinerary (as a starting point) in 5 seconds.
- Getting a quick layman’s understanding of a topic you’re unfamiliar in.
- teens can brainstorm career paths. (tell an llm what they are good at, what they enjoy, and what they dislike)
- need to find a product online (tool, part, etc), but don’t know what it’s called? Describe it to an LLM! Then start your google search journey.
All of that is genuinely useful i think? IMO you’re conflating the _tech_, with how it’s being pushed onto the masses by the tech giants (siri, google search, etc).
Apple recently removed the old Siri backend service and replaced it with "Apple Intelligence", which farms the requests out to whoever Apple is partnering with to provide inference. When this switch happened, it basically broke every user's phone if they relied on Siri.
I'm glad I'm on Android. I can just uninstall Google Assistant, which did the same transformation into a Gemini frontend. I wasn't using Assistant at all, so it isn't a meaningful loss for me.
I read somewhere that they are using tiny models on device for some queries, and then sending requests to apple servers for “complex” things. But i doubt that’s the case tbh. Why? Because even some open source models that i run locally on my mac (deepseek r1) produce better responses than what siri is providing lol. So im very skeptical.
If they truly are processing some siri requests in the cloud, im shocked at how incompetent their implementation is.
i haven’t kept up with siri to be honest. All i know is siri is still working for me, but it’s a degraded experience.
This boring comparison keeps being made but you’re orders of magnitude off in terms of adoption and investment already in a few years time. What a lazy observation parroted by someone that literally can’t see the reality that they live in. To spout off such inane nonsense on hackernews of all places, the place where almost every project is now an AI project is incredible levels of mental gymnastics
Title is mildly misleading, is only available for the API, not the web/mobile interface.
Not sure which title you mean, but the submitted title was "ChatGPT Launches MCP Support" and we changed it to be that of the article (per https://news.ycombinator.com/newsguidelines.html)