Let AI assistants into your product without handing them the keys
How Routine Run's MCP connector lets Claude read runs and act, while every change a model writes waits for a person to approve it on the phone.
On 14 October 2026 Routine Run's MCP connector went into Claude's directory. An assistant can now read your routines and runs, start a run, move it to the next step, and write a new routine for you.
The interesting part isn't what it can do. It's the line between what it does at once and what it only proposes.
Undo decides
Starting, pausing and moving a run happen immediately. Each of those can be undone with one tap, so asking first would only add friction.
Writing a routine, or changing one, is different. The model's version arrives on the phone as a suggestion, shown in full, and the person saves it or throws it away. The tools are even named that way: propose a new routine, propose a change.
The rule fits in one sentence: the machine does what can be undone, the person decides what cannot. It works the same whether the caller is a person, a script or an AI agent.
Read less than you could
Notes and values people type into a run stay hidden from assistants unless the person turns on reading them. An assistant planning your week needs to know which runs happened, not what you wrote in them.
Sign-in without a password
There is no account to sign in to. The assistant shows a code, and you type it on your phone, the way you pair a TV. The phone stays the source of truth; the server in between carries messages and keeps no event bodies.
One thing that will bite people
The MCP spec revision of 28 July 2026 removed sessions: no handshake, no session ID header, any request can land on any server. Good for scaling. But if your tracing groups calls by session ID, that key is gone. Group by the user and the tool call instead, before you upgrade.
If you're building a connector
- Split tools into "acts" and "proposes" by one test: can the person undo it in one tap?
- Expose the least data the assistant needs, and let people opt into more.
- Trace every call with the reason it was made.
What's the first action in your product you'd let an assistant take without asking?
Prema Vision designs and builds AI systems for companies, takes them to production, and keeps a person on them after launch.
More notes
Apple put a language model in every new iPhone. Here's what it's good for
What Apple's on-device model did well in Routine Run, where it failed quietly, and the test that keeps a prompt and its parser from drifting apart.
Our AI interview was losing facts clients gave it. 18 of 19 came back
Three fixes from the Structed AI interview: a safety net that matched the wrong thing, a question asked until the turn limit, and a trace lookup that made replies take a minute.