Building ClawDo — an iOS Task Manager for OpenClaw

OpenClaw crossed 295k GitHub stars in six months. I’ve been running a self-hosted instance for a while now, and at this point it’s the one tool I actually talk to every day – about work, side projects, what needs doing next weekend, what I keep forgetting.

So at some point I noticed something kind of obvious.

My AI already knows what’s on my plate. It’s in every conversation. Every time I said “I should really do X” or “remind me to look into Y” or “we need to fix that thing with Z,” OpenClaw was there. It heard all of it. And every time I had to manually open Todoist and type in the task I’d literally just described to my AI assistant two minutes ago, it felt absurd.

The existing task managers only see what you type into them. That’s a different, much narrower thing.

So I started building something to close that gap. I’m calling it ClawDo. 🦞

The idea: a plugin that lives inside your OpenClaw instance and periodically scans your recent conversations with an LLM. It looks for the clear, actionable things – not vague intentions, not things you’ve already done – and surfaces them as a suggestion queue. You open the iOS app, swipe through what the AI found, accept what matters, reject the rest. The accepted ones become tasks with due dates, reminders, the usual.

The iOS app is native SwiftUI. Today, Upcoming, All Tasks. Local notifications at 9am on due dates. No third-party sync service – it talks directly to your own OpenClaw instance using your existing bearer token. Your conversations stay on your server. Your tasks stay on your server.

The extraction job is where I’ve spent most of my time. Getting an LLM to reliably pull out only the clear, specific, actually-actionable items from a conversation – and not hallucinate tasks you never mentioned, and not resurface things you already rejected – was harder than I expected. The deduplication has a few layers to it. I’m still tuning it. My gate for continuing is simple: if more than 60% of the suggestions turn out to be genuinely useful, the whole thing makes sense. If it’s noisy, the premise falls apart.

The backend plugin is done – SQLite task database, REST API behind the existing OpenClaw gateway auth, the extraction job, a 30-day archive for rejected suggestions in case you change your mind. The SKILL.md format means it should also work with Zeroclaw, Hermes-Agent, and a few others picking up the same plugin interface.

The iOS app is next. I’ve never shipped an iOS app before, so that part will be interesting. :-) I have a plan. We’ll see how it holds up once I’m actually in Xcode. I’m giving myself 45 days before I admit I’m in trouble.

If you’re on OpenClaw and want to try the beta, there’s a waitlist at clawdo.stratinator.com. That’s where the first invites go.

Still very early. The real question isn’t whether this works technically – it clearly will. It’s whether the suggestions are good enough to be worth checking every day. That’s what the beta will answer.

Step by step.

With that – thanks for reading!

Note: English is not my first language. I used AI to help proofread this post – but the ideas, the code, and the bad decisions are mine. 😀