Donely is an AI agent that lives on your Mac. It sees your screen, detects tasks in real-time, and autonomously executes them using your apps—think Jarvis for macOS. It monitors your workflow, answers questions with full context, and completes tasks with one-click approvals.
The New Feature: Remote Computer Use
Today we're shipping remote access. You can now control your Donely-powered Mac from anywhere through our web app. Whether you're traveling or away from your desk, you can kick off automation tasks on your work machine and watch them execute in real-time.
The Technical Stack
Frontend:
- Desktop app: Electron + React + TypeScript
- Web app: React + Vite
- Real-time updates via Server-Sent Events (SSE)
Backend:
- Python FastAPI + LangGraph for AI agent orchestration
- MongoDB + Redis for state management and checkpointing
- AWS Cognito for authentication
The Remote Access Layer:
This is where it gets interesting. We're using Cloudflare Tunnel (cloudflared) to securely expose the local desktop app to the web—no port forwarding, no VPN, no exposed IPs.
Here's the flow:
1. Desktop app spawns a local HTTP server (localhost:8787)
2. cloudflared binary (bundled with the app) creates an encrypted tunnel to Cloudflare's edge network via QUIC protocol
3. Web app connects to your Mac through Cloudflare's global infrastructure
4. All traffic is encrypted end-to-end; the tunnel is authenticated via token
The beauty: it's outbound-only. Your Mac never accepts inbound connections. Cloudflared establishes a persistent outbound connection to Cloudflare's edge, and all remote requests are proxied back through that tunnel.
Why Cloudflare Tunnel?
- Zero trust security: No open ports on user machines
- Global edge network: Sub-100ms latency from anywhere
- Automatic failover: Cloudflared tries multiple Cloudflare POPs if one is down
- Works behind NAT/firewalls: Only requires outbound HTTPS/QUIC
- Built-in DDoS protection: Cloudflare handles all the hard stuff
The AI Stack:
- LangChain/LangGraph for multi-agent orchestration
- Specialized agents: macOS automation, browser automation, web search, personal memory
- Computer Use Automation (CUA) tools: mouse control, keyboard control, screenshots, element detection
The Experience:
Once connected, the web app streams everything happening on your Mac—AI agent thoughts, tool invocations (web searches, app interactions), execution results. You see your Mac's AI agent working in real-time, just like you're sitting in front of it.
What We'd Love Feedback On:
1. Latency: web-based remote control of AI agents—what's acceptable?
2. Use cases: what would you automate remotely that you can't do locally?
This is the first step toward making AI agents truly ambient—they work for you even when you're not around. Next up: multi-device orchestration (coordinate tasks across your Mac, phone, and cloud).
Would love to hear what you think, especially from folks who've built remote access systems or AI agent platforms.
What is Donely?
Donely is an AI agent that lives on your Mac. It sees your screen, detects tasks in real-time, and autonomously executes them using your apps—think Jarvis for macOS. It monitors your workflow, answers questions with full context, and completes tasks with one-click approvals.
The New Feature: Remote Computer Use
Today we're shipping remote access. You can now control your Donely-powered Mac from anywhere through our web app. Whether you're traveling or away from your desk, you can kick off automation tasks on your work machine and watch them execute in real-time.
The Technical Stack
Frontend: - Desktop app: Electron + React + TypeScript - Web app: React + Vite - Real-time updates via Server-Sent Events (SSE)
Backend: - Python FastAPI + LangGraph for AI agent orchestration - MongoDB + Redis for state management and checkpointing - AWS Cognito for authentication
The Remote Access Layer:
This is where it gets interesting. We're using Cloudflare Tunnel (cloudflared) to securely expose the local desktop app to the web—no port forwarding, no VPN, no exposed IPs.
Here's the flow:
1. Desktop app spawns a local HTTP server (localhost:8787) 2. cloudflared binary (bundled with the app) creates an encrypted tunnel to Cloudflare's edge network via QUIC protocol 3. Web app connects to your Mac through Cloudflare's global infrastructure 4. All traffic is encrypted end-to-end; the tunnel is authenticated via token
The beauty: it's outbound-only. Your Mac never accepts inbound connections. Cloudflared establishes a persistent outbound connection to Cloudflare's edge, and all remote requests are proxied back through that tunnel.
Why Cloudflare Tunnel?
- Zero trust security: No open ports on user machines - Global edge network: Sub-100ms latency from anywhere - Automatic failover: Cloudflared tries multiple Cloudflare POPs if one is down - Works behind NAT/firewalls: Only requires outbound HTTPS/QUIC - Built-in DDoS protection: Cloudflare handles all the hard stuff
The AI Stack:
- LangChain/LangGraph for multi-agent orchestration - Specialized agents: macOS automation, browser automation, web search, personal memory - Computer Use Automation (CUA) tools: mouse control, keyboard control, screenshots, element detection
The Experience:
Once connected, the web app streams everything happening on your Mac—AI agent thoughts, tool invocations (web searches, app interactions), execution results. You see your Mac's AI agent working in real-time, just like you're sitting in front of it.
What We'd Love Feedback On:
1. Latency: web-based remote control of AI agents—what's acceptable? 2. Use cases: what would you automate remotely that you can't do locally?
This is the first step toward making AI agents truly ambient—they work for you even when you're not around. Next up: multi-device orchestration (coordinate tasks across your Mac, phone, and cloud).
Would love to hear what you think, especially from folks who've built remote access systems or AI agent platforms.