MCP server built in — connect Claude, Cursor, or any agent→

Pilot your phone fleet.
Or let AI drive.

Helm is a Mac app that mirrors every iPhone and Android on your bench, routes control to your keyboard, and hands the fleet to AI agents over MCP — with human-in-the-loop approvals on every action.

Download for MacWatch the demo

macOS · iOS · Android · MCP-native · Human-in-the-loop

helm — 6 devices connected
MCP :7777
9:41
GmailLive
Pixel 8 · Client Afocused
9:41
ChromeAgent
Pixel 7a · QA-01
9:41
MapsLive
Galaxy S23 · Phoenix
9:41
SafariBeta
iPhone 15 · iOS Labexperimental
9:41
MessagesAgent
Nothing 2 · Dev
9:41
CalendarLive
Galaxy A54 · QA-02

Works with every MCP client

Claude DesktopClaude CodeCursorZedClineContinueWindsurfYour own loop

How it works

From a cardboard box of phones to driven-by-AI in about five minutes.

01

Plug them in

USB in, tiles appear. Helm bundles adb and scrcpy for Android plus the iOS bridge, so there's nothing to install beyond enabling debugging on your phones.

02

Name, focus, drive

Label each phone once. Click a tile to focus it — your keyboard, mouse, and clipboard follow.

03

Hand it to an agent

Paste the MCP snippet into Claude Desktop, Cursor, or your own runner. Approve actions as they stream in.

The console

One pane of glass for every phone on your bench.

Stop juggling scrcpy windows, tmux panes, and a drawer of sticky-note serials. Helm collapses your device wall — iPhones and Android phones alike — into one Mac app that's fast enough to live in.

Mirror every device

Plug them in. Helm auto-discovers each phone, streams its screen into a tile, and keeps twenty devices responsive on a MacBook.

Keyboard-first control

⌘1 through ⌘9 snaps focus between devices. Type, swipe, and paste from your Mac straight into the phone you're driving.

MCP server, built in

Every phone becomes a typed MCP tool — list, screenshot, tap, swipe, type, launch apps, read UI trees. Any agent, any language.

Human-in-the-loop

Reads pass through. Writes prompt. Trust an agent on one device for the next ten minutes — revoke with a keystroke.

Names, not serials

Label each phone once. Helm remembers it across re-plugs, reboots, and machines, so your scripts stop caring about ADB serials.

Activity log, local-first

Every tap, swipe, and keystroke — human or agent — lands in a searchable local log. Replay a session. Blame a bot.

Agents

Hand Claude the wheel. Keep the brake.

Helm ships a local MCP server that any agent can connect to — Claude Desktop, Cursor, or a runner of your own. Read-only calls stream through. Anything that touches a device waits for your OK.

  • Local HTTP at 127.0.0.1:7777, bearer-token auth, never exposed to the network.
  • Streaming HTTP transport — works with every MCP-capable client out of the box.
  • Pause, kill, or rate-limit any agent session from the sidebar.
claude_desktop_config.json
{
  "mcpServers": {
    "helm": {
      "url": "http://127.0.0.1:7777/mcp",
      "headers": {
        "Authorization": "Bearer hlm_***"
      }
    }
  }
}
Exposed tools
list_devicesget_screenshotget_ui_hierarchytapswipetype_textpress_keylaunch_app

Green passes through. Teal waits for your OK.

Also on board

An agent, a scheduler, and a skills library — already inside the app.

You don't need Claude Desktop or Cursor to get started. Helm ships its own chat, its own scheduler, and a place to park the instructions you keep rewriting.

Built-in agent

Chat with your phones. Bring your own keys.

A local chat pane that drives every connected device through Helm's own tools. Anthropic, OpenAI, and Google — whichever API key you paste, that's what it runs on.

  • Claude Opus 4.7, GPT-5, Gemini 2.5 — pick per message.
  • Keys stay on your Mac in a 600-perms settings.json.
  • Reads stream through. Writes wait for your approval.
chat · iphone-15 + pixel-8
claude-opus-4.7
Open the Shopify app on pixel-8 and pull the latest order total.
→ launch_app("com.shopify.mobile") · pixel-8
Today's total is $4,812.30 across 37 orders.
skill: ecommerce-dash
anthropic · openai · google

Schedules

Cron, but for phones.

Tell Helm to warm an account every 20 minutes, poll a dashboard at 9am, or refresh a listing overnight. It ticks through the loop while you're asleep.

  • Interval-based jobs against named devices.
  • Pause, resume, or run-now from the sidebar.
  • Every run lands in the local activity log.
scheduled.json
next tick · 18s
warm-accounts
every 20m
listing-refresh
daily · 04:00
dash-snapshot
every 60m
store-review
weekly · mon

Skills

Save a recipe. Reuse it forever.

Write the instructions once — 'log into this site, accept cookies, export the CSV' — and drop it into any chat or schedule. The skill prepends itself to the system prompt.

  • Named instruction sets stored in skills.json.
  • Pick a skill per chat to shape the agent's behavior.
  • Pair a skill with a schedule for zero-touch runs.
skills
+ new skill
ecommerce-dash

Open the storefront, grab today's orders and revenue.

app-store-diff

Screenshot the listing, diff against last week's capture.

account-warmer

Scroll 8 minutes, like 3 posts, follow 1. Random jitter.

login-checker

Log in on all target devices, flag any 2FA prompts.

Built for

Teams with more phones than hands.

QA teams

Reproduce a bug across fifteen real devices without touching any of them.

Automation engineers

Swap scrcpy plus Python glue for a typed MCP surface any agent can drive.

Mobile growth

Run parallel onboarding, retention, and store-listing tests from one Mac.

Agencies

Manage a rack of client devices and hand each one its own agent and approval policy.

Ecommerce

Monitor listings, pricing, and buyer flows across marketplaces from real phones.

Social & creators

Warm accounts, schedule posts, and farm engagement on real hardware, not emulators.

Research & scraping

Collect data from mobile-only apps through real devices that look like real users.

Solo operators

Keep a box of phones productive. Let Claude do the tapping while you sleep.

Pricing

Start with one phone. Add as many as you can plug in.

Operator

7-day trial
$39/ month

Includes 1 device slot. +$11/mo per additional slot.

  • Scale from one phone to a full bench — iOS and Android, mixed
  • Built-in MCP server for any agent
  • Local activity log + approval controls
  • Free auto-updates through the MVP
Start 7-day trial

No card up front. Cancel from the app.

FAQ

Answers, before you plug anything in.

Do you support iOS?

Experimental

iOS is experimental. Expect rough edges — production-critical work should still run on Android for now.

What works today: device discovery, screen mirroring, taps, swipes, and keyboard input over USB on most iPhones running iOS 16+. What doesn’t (yet): some Apple-protected flows (App Store, certain system settings), background app launches, and a few UI-tree reads. The MCP surface for iOS is a subset of Android’s.

We’re hardening it in the open — feedback from the bench is welcome.

What do I need to run Helm?

A Mac on macOS 13 or later, plus one or more iPhones or Android phones on a USB hub. Helm bundles adb and scrcpy for Android and the iOS device bridge — no manual setup, no Android Studio, no Xcode.

Which agents can drive the phones?

Anything that speaks MCP. Claude Desktop, Cursor, and custom runners all work out of the box. Every connected phone — iOS or Android — shows up as a typed set of MCP tools: list, screenshot, tap, swipe, type, launch, read UI tree.

Where does my data go?

Nowhere you don't send it. Screen streams, taps, and activity logs stay on your Mac. Agents only see what you approve, and revoking a grant is a single keystroke.

How many devices can I connect?

As many as your Mac and USB hub can power. We regularly run twenty phones on a MacBook without dropped frames. Pricing scales per device slot after the first.

Can I try it before paying?

Yes — 7 days, no card up front. Cancel from inside the app. All features unlock during the trial, including the MCP server.

AI × mobile

AI agents, meet your phone fleet.

Helm is the local MCP bridge between Claude, Cursor, or your own runner and every iPhone and Android on your desk. Agents see the phones. You approve what they do. Everything stays on your Mac.

Download for MacRead the docs

Requires macOS 13+ · Android stable · iOS experimental · Windows & Linux later