SiriClaw - an AI agent that executes tasks on your phone.

Inspiration

AI assistants today are intelligent — but passive:

Siri responds.
ChatGPT explains.
OpenClaw automates - but only on a desktop with complex setup.

Yet our real lives happen on our phones. That’s where our messages, emails, calendar, calls, photos, and daily decisions live. We asked a simple question:

What if your iPhone didn’t just answer — but acted?

Instead of manually replying to emails, scheduling meetings, or organizing files, what if you could just say:

“Handle it.”

And it would.

That idea became SiriClaw.

What it does

SiriClaw is an autonomous AI agent for iPhone. It doesn’t just chat - it executes. It can:

Read and reply to emails
Schedule and modify calendar events
Send messages
Make calls
Access files and photos
Trigger workflows across apps
Unlike ChatGPT, it has real access to your phone.
Unlike OpenClaw, it runs directly on mobile.
Unlike Siri, it can autonomously execute multi-step tasks.

You describe what you want. SiriClaw plans and performs it.

How we built it

SiriClaw combines cloud intelligence with local execution. At a high level:

\( Intent by voice command→LLMPlanning→StructuredActions→Control Phone for On−DeviceExecution \)

Wearable device to give voice tasks

We built wearable ring that has a microcontroller, button, microphone and rechargeable battery:

XIAO NRF52840 Sense
Li-Ion rechargeable battery
PushButton
Microphone

The user speeks a natural language instruction.
A cloud-based LLM decomposes the request into structured JSON actions.
LLM (local of cloud-based parses and validates those actions.
A local execution layer triggers native LLM and phone control.

User just need to push the button, tell the task and SiriClaw starts it's magic

Instead of producing text, the model produces executable commands:

{
  "action": "schedule_meeting",
  "time": "3:00 PM",
  "participants": ["john@stanford.edu"]
}

This separation of reasoning (cloud), option to choose local or cloud-based computing and acting (device) allows powerful autonomy while maintaining control.

Challenges we ran into

Access and controlling the phone

Biggest challenge was to find the way how to navigate in phone's menu and "understand" in which state(menu level) is the phone. Most of the time we spent to code that part.

Autonomy vs Safety

Giving AI access to personal data and device controls introduces real risk. We implemented scoped permissions, confirmation layers, and bounded action schemas to reduce unintended behavior.

Translating Language into Reliable Actions

LLMs are probabilistic. Device control must be deterministic. We built structured command parsing, validation pipelines, and fallback loops to improve reliability.

UX Simplicity

Users don’t want configuration panels. We had to design an experience that feels effortless while managing complex automation under the hood. User doesn't need to install any app. You only need to sync via bluetooth and it will start doing its magic.

Accomplishments that we're proud of

Built a working mobile-first autonomous agent
Designed a cloud-to-local execution pipeline
Enabled multi-step task automation on iPhone
Reduced friction compared to desktop-based AI agents
Created a system that acts - not just chats
We moved from “AI assistant” to “AI operator.”

What we learned

Intelligence without execution is limited.
Automation must feel invisible to be powerful.
Safety architecture is as important as capability.
The future of assistants is operational, not conversational.

We learned that autonomy isn’t about flashy AI - it’s about reliable action.

What's next for SiriClaw

Next, we plan to:

Test with users to explore real use-case scenarios where this solution would be important.
Add more actions on the phone for better control
Add more "skills" to Agent.
Strengthen permission transparency and trust
Make it work for Android

Our long-term vision is simple: