Last month a friend who does cross-border e-commerce asked me: Claude already writes his copy, so can it connect to his thirty iPhones and post that copy.
I said yes. He paused, and said he had not expected it this soon.
From 10.6.0 the iOS console supports MCP. WorkBuddy and Claude Desktop can connect to an iPhone and take instructions in plain language. No script to write, no coordinates to tune.
What MCP does for iOS cluster control
MCP is short for Model Context Protocol. It gives an AI model interfaces that can operate the phone.
Before, Claude or WorkBuddy could write copy, analyse data, and suggest ideas, and could not touch the phone. MCP connects that path: the AI uses the console’s interfaces to inject taps, swipes, and input into the iPhone.
One example. A script used to be: open Douyin, tap publish, pick the video, fill the title, add topics, tap send. Six steps, and each step needed coordinates, a wait, and a retry on failure, 45 to 60 lines in all. Now you tell the AI one sentence: post the video on the desktop to Douyin, use the title we settled yesterday, and add the topic #家居好物.
The AI splits that into steps and calls the interfaces one by one.
Lao Chen told me: “I used to write an instruction manual for the phone. Now I tell the AI what I need.”
The difference shows up when you change phone models. Before, every coordinate in the script had to be measured again. One device meant 45 lines changed. Ten different models meant 450 lines. The AI looks at what is on the interface. When the model changes, the pixel positions change, and those coordinates do not need to be redone.
MCP handles decisions and orchestration. Taps, screenshots, and input still go through the iOS automation script interfaces. What it does is turn the flow you described into those calls.
After it is connected, which jobs no longer need a script
Flows that include a judgment pay off first.
The hard part of a script used to be “if this, then that.” A shipping question gets template A, a product question gets template B, a complaint gets a screenshot and goes to a person. The more branches, the longer the script, and one moved button on the platform kills the whole block.
With MCP, those judgments go to the AI. You do not write if-else. You tell it the rules in plain language.
A team that does outsourced customer service, about a dozen accounts, used to sort and reply to DMs with 80 lines of script plus 5 templates. I went through the numbers with them:
| Traditional script | MCP | |
|---|---|---|
| Code | 80 lines + 5 template files | One plain-language rule description (about 200 words) |
| First build | 3 days (write the script + tune coordinates) | 2 hours (write the rules + get a test run through) |
| Upkeep after a platform UI change | 4-6 hours each time (change coordinates + retest) | 0 (the AI recognises the new interface) |
| Time per run | 0.3 seconds | 2-5 seconds |
| Total upkeep per month | 4 hours | 30 minutes |
He said the obvious change was upkeep time. From 4 hours a month down to 30 minutes, 87 percent less. A platform UI change used to cost him half a day of coordinate edits. Now the interface changes and the AI still recognises it. The script layer barely needs to move.
There is also something that used to be hard: working across apps.
Copy a customer message out of WeChat, switch to Taobao, find the matching order, paste the shipping information back as a reply, then switch back to WeChat and confirm. 3 apps, 4 switches, and passing the state through the middle in a script is awkward. My estimate was at least 120 lines of pure script, and every new phone model meant writing it again.
In plain language it is short: take the order number from the message Zhang San sent on WeChat, look up the shipping on Taobao, then send the tracking number and the shipping status back to Zhang San.
The AI can keep the context. The state is still there when it switches apps. A script has to pass that state itself.
What it still does poorly, and when it is worth connecting
The problems I have hit fall into a few kinds.
Latency. MCP goes through AI inference plus interface calls. One operation, from the instruction going out to it running, usually takes 2 to 5 seconds. A traditional script calls the interface directly, in 0.3 seconds. If it has to finish inside a second, stay with a traditional script.
Accuracy. I measured 200 runs. The AI understood the instruction about 94 percent of the time. 6 were wrong, and 4 of those were “tap that red button” landing on the orange one beside it. For operations that need high precision, such as confirming a transfer in a finance app, I still put a human confirmation on the key step.
Network. MCP has to talk to the AI service. If the network drops, it stops. It is a poor fit for an offline setup.
Cost. AI inference is billed by token. I added up the kinds of flows I run day to day:
| Flow type | Tokens per run | Monthly cost at 50 runs a day |
|---|---|---|
| Simple instruction (one step) | about 200 tokens | ¥15-20 |
| Medium flow (3-5 steps) | about 800 tokens | ¥60-80 |
| Complex judgment (with branches) | about 2000 tokens | ¥150-200 |
Running a saved workflow does not count tokens. The first orchestration and the debugging do.
When it is worth connecting: if the flow has more than 3 decision branches, or it crosses more than 2 apps, try it. A fully fixed flow such as open, tap three times, close is faster and steadier as a traditional script.
Then there is the labour cost. Say an operator spends 2 hours a day on branches and exceptions, on a monthly salary of ¥8,000, about ¥46 an hour. 2 hours is ¥92 a day, ¥2,760 a month. After MCP, the operator only handles the exceptions the AI flags, about 20 percent, which saves 1.6 hours a day, worth ¥73 a day, ¥2,190 a month. Subtract the MCP API cost, ¥80 a month at the medium-flow rate, and the net saving is ¥2,110 a month.
API fees, stopping when the network drops, and errors that are hard to trace are all still there. Try one flow first. Add more after it runs. Do not hand every job over at once.
MCP on iOS cluster control is still early. Most people are still working out how to describe a flow clearly in plain language. The toolchain is not mature enough, and when something goes wrong it is hard to trace. I think in another year or two, saying what you need and letting the AI run it will be a common way to work.
Lao Chen eventually connected MCP to the part that bothered him most: customer-service DMs across about a dozen accounts. He did not hand all of it to the AI. About 80 percent of routine replies go through MCP, and the other 20 percent go to a person. What he liked most was that a platform interface change no longer meant rewriting the script.
Lately he has been looking at bringing content distribution in as well.