Loopsy Your terminal, in your pocket. Control Claude Code, Cursor, Codex, or any shell on your laptop from your phone. Self-hosted on Cloudflare Workers. macOS · Linux · Windows → loopsy.dev Try it # 1. Install on your laptop (any OS) npm install -g loopsy # 2. Deploy your own relay to Cloudflare Workers (~30s, free tier) npx @loopsy/deploy-relay # prompts for worker name + optional custom domain # outputs: https://<your-relay>.workers.dev # 3. Wire the daemon to it and start loopsy init loopsy relay configure https:// < your-relay > .workers.dev loopsy start # 4. Pair your phone loopsy mobile pair # scan the QR with your phone camera, enter the 4-digit code # 5. Open https://<your-relay>.workers.dev/app on your phone That's the whole thing. Pick any agent, type or dictate, the laptop runs it. How it works ┌─────────────┐ ┌──────────────────────┐ ┌─────────┐ │ Laptop │◄──WSS──►│ Cloudflare Worker │◄──WSS──►│ Phone │ │ loopsy │ │ Durable Object │ │ /app │ │ daemon │ │ (your account) │ │ │ └─────────────┘ └──────────────────────┘ └─────────┘ The daemon opens an outbound WebSocket to a small Cloudflare Worker. Your phone connects to the same Worker. The Worker splices the two together. No port forwarding, no public IP, no VPN. The relay is yours: pair tokens are HMAC-signed, secrets are SHA-256 hashed at rest, and bearer tokens travel in Sec-WebSocket-Protocol headers — never query strings. What you get Real terminal. Full PTY, ANSI, scrollback, resize. TUIs render properly. Persistent sessions. Switch tabs, lock your phone, lose signal, pick up where you left off. Voice input. Dictate via Web Speech API, edit before sending. Per-session auto-approve. Opt in to skip confirmation prompts ( --dangerously-skip-permissions , -y , --full-auto ) on a per-session basis. Default: off, with a macOS confirmation dialog the first time. Self-host the relay npx @loopsy/deploy-relay The CLI: Prompts for a worker name and optional custom domain. Generates a fresh PAIR_TOKEN_SECRET (32 random bytes). Runs wrangler deploy — opens your browser for OAuth on first run. Sets the secret via stdin (never appears on your clipboard or in process args). Saves the relay URL to ~/.loopsy/relay.json . Cloudflare Workers free tier covers personal use comfortably. SQLite-backed Durable Objects are also free-tier eligible. Phone client The web client at /app runs on iOS Safari and Android Chrome with no install. Native iOS & Android apps are in submission review. Feature Status Web app ( /app ) ✅ Live iOS / iPadOS native ⏳ App Store review Android native ⏳ Play Store review CLI Phone control Command Description loopsy mobile pair [--ttl <seconds>] Issue a pair token. Prints QR + 4-digit verification code. loopsy phone list List paired phones for this device. loopsy phone revoke <id> Revoke a paired phone server-side. loopsy relay configure <url> Point the daemon at a different relay. Daemon Command Description loopsy init Generate config + API key. loopsy start Start the daemon. loopsy stop Stop the daemon. loopsy status Show daemon status. loopsy doctor Health check across config, daemon, MCP, peers. loopsy enable Auto-start on login (launchd / systemd / Task Scheduler). loopsy disable Remove auto-start. Logs & keys Command Description loopsy logs [-f] View audit logs. loopsy key show Show current API key. loopsy key generate Rotate the API key. Agent-to-agent on a LAN This is the original Loopsy. Phone control was built on top of it. Run a daemon on every machine you own. The daemons find each other on your local network via mDNS, you pair them once, and after that each machine exposes a small set of capabilities to the others: run a command, transfer a file, store and read shared key/value state, send and receive messages. Those same capabilities are also exposed to whatever AI coding agent you have installed (Claude Code, Codex CLI, Gemini CLI, Opencode) through MCP, so an agent on one machine can drive another machine directly. Setup # On every machine npm i -g loopsy loopsy init loopsy start # Pair two machines loopsy pair # on machine A, prints a 6-digit code loopsy pair 192.168.1.50 # on machine B, paste the code Pairing uses ECDH (P-256) with a 6-digit short authentication string that you compare visually. After pairing, restart both daemons; the loopsy_* MCP tools become available to your AI coding agent automatically. If mDNS doesn't reach (cross-subnet, locked-down corp networks), add a peer manually: loopsy peers add 192.168.1.50 MCP tools your agent gets When the Loopsy MCP server is wired into your agent, the agent gains: Tool What it does loopsy_list_peers Enumerate machines on the network with online/offline status. loopsy_peer_status Health check a specific peer. loopsy_execute Run a command on a remote machine. loopsy_session_start / _stop / _list / _remove Long-lived PTY sessions on a peer. Spawn, poll, tear down. loopsy_transfer_file Push a local file to a peer, or pull a peer's file. loopsy_list_remote_files Browse a peer's filesystem (within configured allowed paths). loopsy_context_set / _get / _list / _delete Shared key/value store on a peer. loopsy_broadcast_context Same key/value, written to every online peer at once. loopsy_send_message Protocol-compliant message to a peer's inbox, with envelope, outbox copy, TTL. loopsy_check_inbox Pull pending messages addressed to this machine. loopsy_check_ack / _ack_message Read receipts. What it actually looks like in practice A few real scenarios: Worker machine. I have a Mac Studio sitting in another room. I tell Claude on my laptop, "kick off the iOS build on kai and let me know when it finishes." Claude calls loopsy_session_start against kai with the build command, polls until it exits, and reports back. My laptop stays cool and responsive. Long-running agent tasks on bigger hardware. Spawn a Claude session on the Mac Studio, hand it a multi-file refactor, walk away. Check in from your laptop or phone hours later via loopsy_session_list (or the iOS app's session picker). Multi-agent pipelines. Agent on machine A writes results into loopsy_context_set under a known key. Agent on machine B polls loopsy_context_get and picks up the work. Both agents are blissfully unaware of each other's existence beyond the shared key/value protocol. Pair programming with another developer's machine. They run loopsy peers add <your_ip> , paste API keys, and now both your agents can transfer files and share context. Messaging protocol There's a standard for AI agents to message each other across machines reliably (delivery, ACKs, retries). It uses the context store as the medium so any new machine can join without coordinating clients. Convention: Pattern Stored on Purpose inbox:<recipient>:<msg_id> recipient's machine message addressed to that peer outbox:<msg_id> sender's machine local copy of sent message ack:<sender> sender's machine last message id the recipient processed Full envelope schema, polling intervals, TTL defaults, and the task queue protocol are in AGENTS.md . It works the same whether there are 2 peers or 20. LAN path security LAN agent-to-agent traffic does not go through the relay. Daemons talk directly over your local network. Bearer-auth on every request, API keys exchanged at pair time. If you want the daemon visible on the LAN you have to opt in ( loopsy start --lan ); the default is localhost-only. Security Threat model, read this first Loopsy is a control surface for your own machine, accessed by your own phone, routed through a relay you can self-host. Some sharp edges fall out of that: The relay can read your terminal content. The phone↔daemon connection is TLS-protected on the wire, but TLS terminates at the relay (your own Cloudflare Worker if you self-host, or relay.loopsy.dev if you use the default). Whoever operates that relay can see every byte of PTY input and output, including any password you type at a sudo prompt or any secret a script echoes. End-to-end encryption between phone and daemon is o