A week ago I introduced you to Lobster - my personal AI assistant living on a dedicated Mac in my office, accessible via iMessage, managing my inbox, tracking flights, monitoring packages, and generally trying to keep my life a little less chaotic. Today I want to share two things I've been quietly building since then: a public playbook documenting the whole setup, and a brand new project called HomeClaw that gives Lobster the ability to see and control my entire home.

Where We Left Off

When I wrote the original Lobster post, I was still figuring things out. I had something working — Lobster could triage my email, respond to texts, check on packages, manage my calendar — but it felt rough around the edges. More proof-of-concept than finished product.

What surprised me was the response. A lot of people reached out asking the same questions: Are you crazy? This can’t be secure. Wow! How do I get one!

Those questions became the seed for the playbook.

The Playbook: My Setup, Documented

I’ve been quietly writing down everything at lobster.shahine.com.

It covers my full OpenClaw setup from the beginning: the hardware, the iMessage bridge, the MCP servers, the multi-agent security architecture, email handling, remote access via Tailscale, and everything I’ve added or changed since launch. It’s less of a tutorial and more of a living notebook — the decisions I made, the tradeoffs I accepted, and the things I’d do differently starting over today.

The changelog captures what’s changed week by week. Since Lobster’s initial release on February 16th, a lot has moved: FlightRadar24 real-time flight tracking landed on the 17th, the documentation site went live on the 18th, native Apple PIM plugins replaced a tangle of MCP wrappers on the 19th (cutting ~5K tokens per context window), the BlueBubbles Private API unlocked typing indicators and read receipts on the 21st, and voice mode via ElevenLabs and the OpenClaw iOS app went live the same day. It’s been a busy week and a half.

What Lobster actually does all day

Before I get to the new stuff, let me show you what the last few weeks have looked like in practice. Every one of these is real — pulled from actual message logs.

“Tell Omar the kids want to see a movie”

My wife is at the hotel with the kids. I’m not looking at my phone. She texts Lobster: “Hi Lobster can you text Omar to let him know that the kids want to see a movie today — 12:30 showing at the theater nearby.” Lobster relays it to me. She follows up: “What time do you recommend we leave the resort to park, get refreshments, and make a 12:30 movie?”Lobster searches drive times, adds a buffer for parking and snacks, and replies: leave by 11:45 AM. No app switching. No Googling. Just an answer.

“Where did we travel last year?”

My daughter drops a question in the family group chat: “Lobster, do you know where we traveled last year?” Lobster checks the family Travel Hub (a TripIt I built for myself), finds the trip — dates, duration, nights. My daughter follows up from memory: “I know we stayed at a nice hotel on the water near a famous landmark.” Lobster identifies the hotel. Suddenly the whole family is reminiscing in the group chat, with Lobster filling in the details nobody could quite remember. She then asks: “What are our next two trips? Where are we staying?” Trip planning feels less like work and more like anticipation.

Your flight just landed

You didn’t ask for this. Lobster was already tracking the flight via FlightRadar24. When the wheels touch down, you get a landing notification, check-in instructions for the hotel, driving directions from the airport, and a reminder with arrival details. Your spouse gets the same info in the family group chat — directions, tipping guides, what to expect at check-in.

You never missed that 7 AM meeting

Lobster checks the calendar twice a day and knows the patterns: early morning meetings (7–8 AM) get a reminder the evening before; late meetings (after 8 PM) get one at 5 PM day-of. You didn’t set these up. You didn’t configure rules. Lobster just pays attention.

“Play Bad Bunny on Sonos”

“Play Bad Bunny on Main Floor Sonos.” Music starts. Multi-room Sonos, Spotify and Apple Music, per-room volume defaults — all from a text message.

These aren’t edge cases or demos. They’re Tuesday. The full use cases page — including how the family’s three-agent permission model works, and what the 30-minute heartbeat monitors — is at lobster.shahine.com/use-cases.

HomeClaw: Lobster Comes Home

Here’s the new thing I’m genuinely excited about.

Lobster has always been great at managing information — email, calendar, travel, packages. But there was always this gap: Lobster couldn’t see or interact with the physical house. My lights, thermostats, locks, sensors, speakers — all of that lived in Apple Home, siloed behind the Home app, invisible to my assistant.

HomeClaw closes that gap.

It’s a Mac app I built using Claude Code (in about 6 hours) that bridges Apple Home / HomeKit to Lobster — and to anything else in the OpenClaw and Claude Code ecosystem. Under the hood, it exposes both an MCP server and a CLI, which means:

• Lobster can now query and control HomeKit accessories as naturally as it handles your calendar

• Claude Code can use the same interface if you want to write scripts or automations

• Any MCP compatible agent can tap into your home through a clean, consistent API

In practice, this means I can say “Turn off everything downstairs and set the thermostat to 68 for the night” over iMessage and Lobster just handles it. Or I can ask “What’s the temperature in the kids’ room?” and get an actual answer. The home stops being a black box.

Here is a good example. “Is Miles awake”?

Answer: “His overhead light is off and the Nanoleaf is off. The presence sensor in his room is unreachable so I can’t check motion. Based on lights alone, looks like he’s either asleep or not in his room -- but it’s 10:40 PM on a school night so hopefully asleep!”

Can Siri do that? Nope. She can barely tell me the time.

I wrote HomeClaw because I couldn’t find anything that connected HomeKit to the agentic AI world in a way that felt native. Most solutions are either overly complex home automation scripts, or cloud-dependent black boxes that route your home data through someone else’s servers. HomeClaw runs locally, talks directly to HomeKit via the Mac APIs, and exposes a standard interface that fits naturally into the tool-calling model these agents already understand.

What I’ve Learned Running This for Real

A few things I didn’t expect when I started:

The mundane stuff is where the value is. I thought I’d be most impressed by complex multi-step tasks. What actually moved the needle was reliability in the small things: Lobster noticing an early meeting and sending a reminder the night before without being asked. That kind of invisible helpfulness builds a lot of trust, fast.

Context compounds. The more Lobster knows about the family — preferences, patterns, past trips, who we talk to — the more useful it becomes. The implication is that building a good assistant is a long-term investment, not a one-time setup. The first week is a rough draft. The second week is noticeably better.

Privacy architecture matters from day one. Because Lobster runs on hardware I own, my family’s calendar, messages, and now home data stay on our network. The more capable the assistant gets, the more important it becomes to trust where your data is going. Local-first is a feature, not a constraint.

The interface you don’t think about is the right one. iMessage turned out to be perfect — it’s already where the family communicates, it requires zero behavior change, and it works the same whether you’re home or traveling. The best interface is the one everyone already uses.

What’s Next

HomeClaw is early. It works well for my setup and if I can figure it out I’ll publish to the App Store, but for now if you have an Apple Developer account you can build and install it on your Mac.

The playbook is a living document. I’ll keep updating it as the setup evolves — new MCP servers I’ve added, things I’ve changed, integrations that didn’t work out.

If you’re building your own assistant, have questions about the setup, or want to contribute to HomeClaw, I’d love to hear from you. The links are below.

— Omar

Everything I’ve Built for Lobster

All of these are open source. Use them, fork them, build on them.

🦞 HomeClaw — HomeKit smart home control via MCP and CLI. Control lights, locks, thermostats, and scenes from Claude Desktop, Claude Code, or OpenClaw. Native macOS menu bar app with a split-process architecture to work around Apple’s HomeKit restrictions.

🍎 Apple PIM — Native macOS integration for Calendar, Reminders, Contacts, and Mail using EventKit and Contacts frameworks. Works with Claude Code (MCP) and OpenClaw.

🔌 Agent Plugins — Plugin marketplace for Claude Code. Includes Chief-of-Staff (email triage & productivity orchestrator), travel agent, credit card benefit tracking, and AI-powered file renaming).

📧 Fastmail MCP — Remote MCP server for Fastmail running on Cloudflare Workers, secured with Cloudflare Access OAuth.

✈️ Travel-Hub — My Private Hub for all Flights, Hotels, Car Reservations and more. It’s crazy awesome but it’s just for my family.

Links:

• 🌐 Playbook: lobster.shahine.com

• 🔧 HomeClaw on GitHub: github.com/omarshahine/HomeClaw

• 📬 Original post: Meet Lobster 🦞: My Personal AI Assistant