The Laptop is Home. You're on a Train. Your AI is Working.
Anthropic's March 2026 update brought three distinct ways to control AI agents from your pocket. Stop the LinkedIn hype-scroll and learn what actually works.
The LinkedIn post everyone is sharing is half right. The idea is real — you can run and manage AI agents straight from your phone, no OpenClaw, no VPS, no reverse proxy. The catch: what people call "Claude Dispatch" is actually three separate features, each solving a different problem. Confuse them and you set up the wrong thing.
There isn't just one 'Claude Dispatch'. Anthropic shipped Cowork Dispatch (office tasks), Remote Control (developers), and Channels (Telegram/Discord integration).
Which Feature Do You Need?
The Mobile Claude Ecosystem
- User: Knowledge workers.
- Goal: Assign desktop tasks (email drafting, file organization) from your phone.
- Requires: macOS, Pro/Max plan.
- User: Software developers.
- Goal: Steer active coding sessions, fix CI/CD failures, trigger builds.
- Requires: Claude Code v2.1.51+.
- User: Automation power users.
- Goal: Two-way Telegram/Discord bots with access to your local machine's context.
- Requires: Claude Code v2.1.80+.
Security Architecture: The Secure Relay Pattern
Unlike typical remote-desktop tools, Claude Code never opens inbound ports (no SSH exposure). It uses a relay layer to keep your machine hidden.
- Machine: Your local terminal runs Claude Code and initiates an outbound connection.
- Relay: Anthropic's secure relay layer handles the handshake.
- Phone: You connect to the relay via a short-lived, encrypted OAuth session.
- Result: Zero open ports, 10-minute automatic timeout on inactivity.
When Claude uses computer control, it can be 'prompt injected' by malicious content on your screen. Never run Dispatch against untrusted documents, public URLs, or open browser tabs with third-party sites.
Part 1: Cowork Dispatch — The "Executive Assistant" Mode
Built into the Claude macOS app, Dispatch delegates "Computer Use" tasks from your iPhone. It's an assignment tool, not a remote desktop — you assign a task and check back for the report.
Cowork Workflow
Open Claude on Mac → Settings → enable 'Remote Tasks'.
On iPhone: "Find the invoice in my Downloads, summarize the totals, and draft an email to accounting."
Claude takes control of your screen — you see the 'hand' icon on your Mac moving.
Task finishes; you get a push notification with the draft or report.
Honest caveat: computer-use success on real tasks is ~50% today and macOS-only. It improves as Claude's computer-use capability improves.
Part 2: Remote Control — Steer Coding Sessions From Anywhere
The feature most developers actually want. Run Claude Code, step away, check in and steer from your phone.
claude --version # need v2.1.51+
npm update -g @anthropic-ai/claude-code
# Method 1: fresh remote-enabled session (prints a QR code + URL)
claude --remote-control "my-project-name"
# Method 2: enable on an existing session
/remote-control my-project-name # spacebar toggles the QR displayThe machine never becomes a server: Machine → Anthropic relay → phone/browser, outbound HTTPS only, short-lived single-purpose credentials, hard 10-minute network timeout. Requires a paid claude.ai OAuth login (API keys don't work; Bedrock/Vertex/Azure not supported).
From your phone you can: watch Claude Code work in real time, send new instructions mid-task, approve permission requests, review changes, and see errors as they happen. You can't: keep the session alive if the terminal closes or the machine sleeps.
Never combine --dangerously-skip-permissions with --remote-control. That gives remote access with no permission gates — anyone with the session URL gets unsupervised terminal access. Keep gates on; approve each request from your phone.
Part 3: Channels — Telegram & Discord Inside Your Sessions
The most experimental feature, and the one that answers the "OpenClaw killed" framing. Bidirectional: Claude Code messages you at milestones and asks for decisions; you reply with instructions or approvals. Telegram and Discord only, in the March 2026 research preview.
# 1. Create a bot with BotFather on Telegram → get BOT_TOKEN
# 2. Update + verify
npm update -g @anthropic-ai/claude-code # need v2.1.80+
# 3. Start a session wired to the channel
claude --channels plugin:telegram@claude-plugins-officialFor Discord, create an application in the Discord Developer Portal, add a bot, and use the Discord channel plugin instead. Once running, message your bot: "What's the status of the build?", "Deploy the latest commit", "Find all TODO comments in the auth directory."
Even on mobile, Claude asks before running dangerous commands (e.g. rm -rf). You approve the action directly in the chat window.
Real-World Workflows: What to Actually Build
- The long build monitor (Remote Control).
claude --remote-control "build-monitor", then: "Run the full test suite. Update me when a section passes/fails. If anything fails 3× in a row, stop and tell me exactly what happened." Check in from your phone every 20 min. - The document assistant (Dispatch). From your phone: "Open each file in Desktop/Q1-Reports, extract key metrics, build Q1-Summary.md with a table, email it to me." Your Mac handles it.
- The background task with a decision point (Channels). "Run /scripts/migrate.py. After 500 rows, Telegram me a 5-row sample to verify the format. Wait for my reply before continuing; send a final count when done."
- Parallel research (Remote Control ×2). Two sessions (
research-session-a&-b) running at once; steer whichever needs attention from your phone.
Anthropic vs. OpenClaw: The Honest Truth
The Agent War
- Pro: SOC 2 Type 2 + ISO 27001, 1-command setup, seamless mobile apps.
- Con: macOS-heavy, requires a $20/mo+ subscription, Claude-only.
- Pro: Free, 20+ platforms (WhatsApp, Signal, Slack, iMessage), full Linux, multi-model, scheduled HEARTBEAT tasks.
- Con: Hard setup (VPS, SSH tunnel, reverse proxy), 288 security notices, ClawHavoc malware incident.
What Anthropic got right: setup simplicity (Remote Control is one command vs. hours of OpenClaw config), security infrastructure (compliance vs. OpenClaw's Atomic Stealer / keylogger incidents), and a genuinely excellent developer workflow. What OpenClaw still does better: platform reach (20+ vs. 2), scheduled proactive tasks (the viral "7am priority summary" use case Anthropic doesn't replicate), and multi-model + free. Conclusion: Anthropic matched the headline feature, made it far easier and more secure; OpenClaw keeps the edge for always-on, broad-platform, multi-model power users.
Troubleshooting & Where It's Headed
Common issues: wrong/old version (claude --version then npm update -g @anthropic-ai/claude-code), API key instead of OAuth (Remote Control/Channels require claude.ai login), session dying when the terminal closes (use caffeinate on Mac to keep the machine awake). Roadmap: Dispatch → Windows support + better computer-use accuracy; Remote Control (stable/GA) → persistent sessions surviving terminal restarts; Channels → WhatsApp & Slack most-requested, no committed timeline.
Key Takeaways
Dispatch is an assignment tool. You don't watch it work; you assign a task and check back for the report.
Never use --dangerously-skip-permissions with Remote Control. Always approve tool use from your phone to stay in control of your filesystem.
Sessions die if your terminal closes or machine sleeps. Use the caffeinate utility on Mac to keep tasks running while you're away.