The personal assistant shift is the most interesting part of this. I've seen the same pattern with Claude Code: once you give it access to communication tools, coding becomes a fraction of what you use it for.
One thing we ran into building tools in this space (I work at Nylas on email/calendar infrastructure): depending on how Claude Code accesses your mail and calendar, things get fragile fast. Weird message formats, long threads, complex recurrence rules on calendar invites. We built a CLI (nylas) that gives agents structured email/calendar access from the terminal specifically because the ad-hoc access paths kept falling over for anything beyond simple reads.
Curious how you handle the email/calendar side in practice. Is Claude Code going through Apple Mail directly, or do you have MCP servers wired up for that?
The single API call to get an email address up and running is a nice DX touch. We've seen at Nylas how much friction OAuth flows and domain verification add when all you want is programmatic email access.
Curious about deliverability on the shared domain — have you run into issues with @agent.mailboxkit.com landing in spam? In our experience, shared sending domains hit reputation problems fast once multiple tenants are sending from them, especially for outreach use cases.
The LLM-readable skill file idea is clever. We've been doing something similar with llms.txt for AI discoverability.
We also just shipped an OpenClaw plugin for Nylas (https://www.npmjs.com/package/@nylas/openclaw-nylas-plugin) that gives agents email, calendar, and contacts access out of the box — different approach since it connects to real user mailboxes rather than creating new addresses, but solves the same "agents need email" problem.
One thing we ran into building tools in this space (I work at Nylas on email/calendar infrastructure): depending on how Claude Code accesses your mail and calendar, things get fragile fast. Weird message formats, long threads, complex recurrence rules on calendar invites. We built a CLI (nylas) that gives agents structured email/calendar access from the terminal specifically because the ad-hoc access paths kept falling over for anything beyond simple reads.
Curious how you handle the email/calendar side in practice. Is Claude Code going through Apple Mail directly, or do you have MCP servers wired up for that?