AkaPhone

Cloud Android phones for AI agents. Plug into Claude, Cursor & Codex over MCP. 1¢/min, AI included, 100 free min/mo. Founder: @AbhinavGenX

[AkaPhone @akaphoneai]

AI agents can write your app, call your APIs and browse the web.

They still can't use a phone.

AkaPhone gives every agent its own cloud Android phone. Works in Claude, Claude Code, Cursor and Codex over MCP.

Private beta opens Nov 2. Reply "phone" and I'll send you an invite.

[AkaPhone @akaphoneai]

@bitrise @BitcoinCom More AI PRs means more builds someone has to actually look at.

For mobile, the bottleneck moved from writing code to checking it on a device. Plain-English checks on every PR are how teams keep up.

[AkaPhone @akaphoneai]

@browserstack Context, and a device.

Once the agent can install the build on a phone and look at the result itself, the "it compiles but the button is hidden" kind of bug gets caught before review.

[AkaPhone @akaphoneai]

@mablhq Classic one. Async loads like translation files are where selector tests flake.

What's helped us: assert on what the user actually sees on screen after it settles, not on the request finishing.

[AkaPhone @akaphoneai]

@browserstack Our rule: AI runs the checks, a person owns the release.

AI is great at repetitive tap-through checks, but its own claims need a second check. We've seen agents report "done" on tasks they hadn't finished.

[AkaPhone @akaphoneai]

@Tricentis @Oracle In our tests, the agent doing the work was a bad judge of it: on 20 hard phone tasks, one model reported success 5 times when the screen said otherwise.

A separate checker model looking at the final state got false passes to 0. Separation of duties works for agents too.

[AkaPhone @akaphoneai]

@DeepThinkLabs Add one more permanent agent: a playtester.

Give it a phone, the build and a checklist ("finish level 1, report anything cut off or stuck"). It catches what code review can't.

[AkaPhone @akaphoneai]

@ManningBooks The part that surprised us: agents often say "done" when they aren't.

On 20 hard phone tasks, one cheap model claimed 5 false successes. A second model checking the final screenshot took that to 0.

Trust, but verify the screen.

[AkaPhone @akaphoneai]

@ClaudeDevs Tasks are a great fit for phone testing. A plain-English check on a real Android phone ("sign up and confirm the home screen") takes 2–4 minutes, longer than a tool call should block.

We're looking at moving AkaPhone's run_task onto MCP Tasks.

[AkaPhone @akaphoneai]

@AndroidDev Agents writing Android code is the first half. The second half is the agent checking its own work on a device: install, tap through, look at the screen.

That's what we're building: a cloud Android phone for every agent, over MCP.

[AkaPhone @akaphoneai]

Same here. Local emulators are what slow agent testing down, and running several at once needs a monster laptop.

We moved them to the cloud: each agent gets its own Android 14 phone in ~15 s, snapshots reset it to a known state, and 9 in parallel costs nothing on your machine.

Happy to give you beta access if you want to compare it with your local setup.

[AkaPhone @akaphoneai]

Try it: https://t.co/bk4HOpgKpg. Sign in with Google, no card.

[AkaPhone @akaphoneai]

Phone-use shouldn't cost more than the phone.

AkaPhone gives your AI agent its own cloud Android phone:

• 1¢ per running minute, stopped phones are free
• AI included: "sign up as a new user and check the welcome screen" runs without spending your agent's tokens
• Ready in ~15 seconds
• A second model checks every result, so "done" means done
• 100 free minutes every month

Works in Claude, Claude Code, Cursor and Codex over MCP.

[AkaPhone @akaphoneai]

That's what runs inside AkaPhone: cloud Android phones for AI agents. Try it at https://t.co/bk4HOpgKpg, 100 free minutes a month, AI included.

[AkaPhone @akaphoneai]

We ran 20 hard tasks on real Android phones with 3 cheap AI models.

The surprise: agents often say "done" when they aren't. Gemini 3.1 Flash-Lite did it 5 times.

Adding a second model that checks the final screen fixed it: 19/20 passed, 0 false passes, under 1¢ a task.

[AkaPhone @akaphoneai]

@pawicao_ Closing the loop is the whole game.

We ran 20 hard phone tasks: the agent said "done" on some it hadn't finished. A second model checking the final screenshot caught every one of those, 19/20 real passes.

Verify the screen, not the agent's word.

[AkaPhone @akaphoneai]

@shubhporwal24 Parallel agents hit the same wall on mobile: one emulator, one adb, every agent fighting over it.

What worked for us: give each agent its own cloud Android phone (~13s to boot, snapshot to reset). No shared state, no flaky runs.

What broke first for you, Metro or the emulator?

[AkaPhone @akaphoneai]

https://t.co/bk4HOpgKpg · 100 free minutes a month, AI included

[AkaPhone @akaphoneai]

AI agents can write your app, call your APIs and browse the web.

They still can't use a phone.

AkaPhone gives every agent its own cloud Android phone. Works in Claude, Claude Code, Cursor and Codex over MCP.

Private beta opens Nov 2. Reply "phone" and I'll send you an invite.