> DoorDash only rolled out its Model Context Protocol (MCP) connector less than a month ago. ... Without that kind of linkup—which is still only available to a small number of “early beta partners” and mainly used to streamline corporate lunch orders

So it's available to almost no one anyway. Misleading title.

I would not trust the AI to order food off of these websites for me because the labeled data itself is pretty much shit. Lots of dishes are unlabeled, have really bad pictures or no pictures at all, and the worse is the ones that have AI generated descriptions which are often wrong or not helpful. I'm totally fine with letting it just use my saved credit cards and delivery address (I assume the AI doesn't see them) but I can't trust it to order the food I actually want.

DoorDash seems like a resource anti-optimization that would be made irrelevant by an AI agent managing one's meals. An unofficial HEB MCP [0]—that's cooking with gas!

0. https://github.com/mgwalkerjr95/texas-grocery-mcp

* No affiliation with the above repo, but I'm considering trying it out.

Because the DoorDash app is so simple, delightful, intuitive?

Look, there's a lot of AI excitement and a lot of it is dumb. But if there's a number I can text "hey get me a pizza" and it re-orders the same pizza order I make every Friday night, I'm going to use it. Would I pay for it? As a standalone service? Obviously not.

But if I could delete half the apps on my phone with their horrid dark-pattern-filled interfaces and constant upsells, I'd love that. I'll happily delegate the interfacing to an agent.

And I'd 100% pay that agent to drop the ads and upsells when DoorDash tells it to pass them along.

Remember when Amazon thought that people would say "buy me more paper towels" to their Echo? And of course they didn't because that's insane.

Something that is perhaps “underreported” is that a lot of AI solutions aren’t easier to do than just doing the thing.

My budget app, for example: It’s not easier for me to ask the AI bot “show me all my transactions from McDonald’s” than it is to click on the search bar/payee filter and type “McDonald’s.”

But if I use the AI bot, I’m burning tokens.

2026, re-inventing APIs with 3 levels more indirection \_(ツ)_/ and calling it "innovative"

This is kind of a moot article because up until a month a go when DoorDash released their MCP, their bot detection was getting in the way of agentic ordering.

Some of these use case feel so imaginary that lot of people just don't get why would some other people do them. Like with food delivery I want to select specifically what I want and evaluate if the price is right. Same goes for ordering most stuff or deciding on vacation.

Maybe I could get train ticket from AI... Or order taxi. At least until it went wrong for the first time...

For me, the best use of AI is making bespoke software connecting systems and devices that have no business being connected but I find value in so I can have a unique interface for all my tech and hobbies. Not sure how that gets translated to the everyday man but it sure isn't "bring me food".

Agents are good for complex purchases with lots of hidden fees like hotels. Less useful for burgers, though I'd definitely try it for a large catering order.

When I have demoed the use of agents, I usually do it with an agent combining multiple applications' results. Something like: "We're having a family gathering at grandma's. Which of us can stop in at the drugstore to pick up her prescriptions on the way without adding a lot to our travel time?"

Agents are great at moderately complex combinations of data and apps. If I had to do this every day or every week even, it might be worth creating a script suited for a task like that using some automation system. Agents are great at providing a disposable saw when it doesn't make sense to stop and sharpen the saw.

I'm not a doorDash user, but they've been doing it long enough to have an app that's probably highly polished. There's not going to be a lot of upside to speaking or typing an English language command to get what you want. Trying to do that is a waste of a well designed app and the agent's token budget.

Bots and tech are testing humanity's levels of desire for laziness and how much they are ready to pay for buying that laziness. Turns out that the desire is limitless. As someone said, the ultimate luxury is in having maximum laziness - not having to move, talk, think or use senses - the state of being a rock.

Not surprising in hindsight.

It's not a super repetitive problem (e.g.respond to 100s of messages) and human judgment is highly valued here (what am I hungry for?) and visuals are sometimes a factor.

The problem with AI assistants in general is that humans barely ever know what they want and the promise is that AI can intuit/predict what we want. It can't. People like to shop. It's easier to shop with our eyes than our ears. Personal assistants, as far as I can tell, have one use case and that is doing mundane chores like e-mail summaries. But to do that, you have to give them access to everything, so it becomes a transaction of privacy. And I think a lot of people are paranoid. Plus each one is a little token bonfire that is probably costing them money.

And I feel like the labs must all know this, so it makes me wonder why they're actually trying to push these assistants onto us. The answer I'm left with is training data — data theft. They'd all violate their privacy terms if it earned them an extra nickel.

I know people that have been automating their Doordash orders even before modern AI was a thing. When done right the server cannot tell.