Phonebox: Cloud Android Phones for AI Agents, Explained
Phonebox gives AI agents an Android phone in the cloud. Your agent, whether Claude Code, Codex, Cursor or something you built, creates a phone with one call, reads its screen as a short numbered list of elements, then taps, types, swipes, installs apps and moves files. When the agent is done, it “parks” the phone, which stops billing while keeping the apps, files and signed-in accounts for next time. Phones cost $0.06 a minute while running, billed by the second. Phonebox is operated by Steelworks Software LLC.
The quick answer: Phonebox is a focused, developer-first take on phones for agents: Android only, metered by the second, and documented almost entirely with coding agents in mind. Its sweet spot is letting a coding agent install and test the Android app it just built. It’s young, and we found little independent coverage of it. A disclosure: we make Carly, an AI assistant that makes and answers phone calls and texts, which Phonebox says plainly it doesn’t do. Everything below comes from Phonebox’s site, docs and terms.
Phonebox at a glance
Checked against Phonebox’s site, documentation and terms on November 12, 2026.
| Phonebox | |
|---|---|
| What it is | Persistent cloud Android phones that AI agents see and control |
| How agents connect | A hosted MCP server, a CLI, a TypeScript SDK, a REST API and an agent skill file |
| Works with | Claude Code, Codex, Cursor, Gemini CLI, Windsurf, Cline, GitHub Copilot and any MCP client |
| Devices | Virtual Android phones, which can appear in 48 countries |
| Price | $0.06 per phone-minute, billed per second; plans from $20 a month |
| Free trial | 7 days with $5 of phone time, card required |
| Company | Steelworks Software LLC, Delaware |
What Phonebox does
The docs put the whole product in two sentences: Phonebox “gives your agent an Android phone in the cloud: its screen, its installed apps, touch input and the keyboard.” And just as clearly, “Phonebox is not a phone number or an SMS inbox, and it doesn’t make or receive calls. Your agent controls the device itself.”
It names three audiences:
- Coding agents testing Android builds. Claude Code, Codex or Cursor builds an APK, installs it with one command, and walks through the screens it changed.
- Agents using apps with no API. When an app’s screen is the only way in, an agent can operate it directly.
- Teams building agent products. A company can give each of its customers their own phone, tagged with a customer ID and reachable only by keys scoped to that customer.
Phonebox’s guides lean heavily toward testing, with walkthroughs on reproducing bug reports, checking deep links and push notifications, accessibility audits, location-based apps, and running Android checks in GitHub Actions without an emulator.
How an agent drives a phone
The work happens in a loop: the agent reads the screen, acts on it, and reads it again to see what changed.
Reading the screen. An observation returns which app is in front, whether the keyboard is showing, and every element the agent can act on or read, each with a numbered reference. The agent can ask for compact text instead of full JSON, plus an optional screenshot. Reading the screen as text rather than pixels keeps each step small for the model.
Acting. Actions are sent in batches, such as open an app, wait for the “Sign in” button, tap it, type an email address and submit. Targets can be an element’s text, ID or reference number, so the agent doesn’t have to work out coordinates.
Device settings. The agent can set the phone’s GPS location, locale and time zone, record the screen, set the clipboard, and reboot or reset the phone.
Interfaces. Every interface calls the same service. The hosted MCP server lives at phonebox.dev/mcp, the API at phonebox.dev/v1, and the CLI and TypeScript SDK ship in one npm package. Phonebox also publishes every docs page in Markdown, an llms.txt index and an agent skill file, so a coding agent can learn the whole workflow from one link. Setup can be done entirely by the agent: you paste one prompt, and you only sign in and start the free week from a link.
Parking and persistence
Phonebox’s central idea is that a phone persists while costing nothing when idle. “A parked phone keeps its apps, files and signed-in accounts, and it costs nothing” until it starts again.
Phones park in three ways: the agent parks them when it’s done, a phone that nothing has addressed for a set idle time parks itself (five minutes by default, up to an hour), and every session has a maximum length, 15 minutes by default and up to three hours. Phonebox’s rules for agents are blunt: “Always park a phone when you’re done, even when the task failed. A phone bills from the moment it exists until it parks.”
This makes Phonebox practical for agents that come back to the same signed-in apps day after day, without paying for a phone to sit running.
Signing in and handing off to a person
Many apps need a sign-in before an agent can do anything. Phonebox’s sign-in guide has the agent drive most of it, then hand the phone to a person when a step needs them, such as a code sent to their own phone.
The handoff is a live view link. The agent creates it, sends it to you with plain instructions like “Enter the code we sent you, then tap Verify,” and waits for the next screen. Phonebox is careful about two things here:
- Typed text isn’t stored. “Phonebox never stores the text your agent types.” Activity logs record it only as “[redacted]”.
- Live links are powerful. “The link lets anyone who opens it control the phone, and every account signed in on it, until it expires.” Send it privately, with the shortest expiry that works.
The guide also sets the ground rule: “Sign in only to accounts that your user owns or may use, and only as the app’s own terms allow.”
Pricing and plans
A phone costs $0.06 a minute while it’s starting or running, which works out to $0.001 a second, with a 60-second minimum per session. Parked phones cost nothing, and a session that fails before the phone is ready isn’t charged. When a session starts, Phonebox reserves credit for its maximum length, such as $0.90 for the default 15 minutes, and releases whatever isn’t used.
Phone time comes through a plan:
| Starter | Pro | Enterprise | |
|---|---|---|---|
| Price | $20 a month | $99 a month | Talk to sales |
| Phone time included | $20 | $99 | Custom |
| Phones at once | 5 | 20 | Dedicated |
| Saved phones | 10 | 50 | Custom |
| Recordings | 20, kept 30 days | 100, kept 90 days | Custom |
| Support | Email, 1 business day | Shared Slack, same day | Named contact and SLA |
Each plan’s price comes back as phone credit, and credit you’ve paid for doesn’t expire. New accounts get a free week with $5 of phone time and Starter’s limits. A card is required, and nothing is charged for seven days. Every project has a monthly spending limit you can change, and a new session won’t start if it would push the project past it.
Rules on what you can do with it
Phonebox’s terms allow use “only with apps, accounts and content you are authorized to use,” and list what’s prohibited, including fraud, spam or deceptive automation, “creating accounts at scale against a service’s terms,” and “circumventing the security measures or rate limits of third-party services, for example through credential stuffing or CAPTCHA farms.” Phonebox can suspend phones that break the rules and freeze the project while it investigates.
The terms are also honest about the hardware: Phonebox phones “are virtual Android devices running in cloud data centers. Some apps behave differently on them than on a handset, and some may not run at all.” Capacity is limited, and there’s no uptime commitment unless agreed in writing.
How Phonebox compares
Phonebox is one of several services giving agents phones. PhoneUse also rents cloud Android phones by the minute and adds robot-tapped real iPhones. Glasshand rents dedicated physical iPhones and Macs by the month. TapKit turns your own iPhone into a phone your agent can drive. Droidrun, now Mobilerun, pairs an open-source framework with a cloud that includes real iPhones, phone numbers and email inboxes. Phonebox is the narrowest of them: Android only, no phone numbers, metered by the second, with the most attention paid to coding agents testing their own builds.
If what you need is phone calls and texts rather than an agent working through apps, Carly does phone and SMS out of the box: it makes and answers calls and texts from a real business number, and handles email and calendars too, from inside ChatGPT or Claude or by text.
Who Phonebox suits
Phonebox is a good fit if:
- You want Claude Code, Codex or Cursor to test the Android app it just built, without running an emulator.
- You want agents to use Android apps that have no API.
- You’re building an agent product and want one persistent phone per customer.
- You’d rather pay by the second than for a phone that sits idle.
Look elsewhere if:
- You need iPhones or real physical devices.
- You need the phone to have a number for texts or calls.
- You need an uptime commitment without an enterprise agreement.
- Your plan involves creating accounts at scale or getting around another service’s protections, which the terms forbid.
FAQ
What is Phonebox?
Phonebox is a service from Steelworks Software LLC that gives AI agents persistent cloud Android phones. Agents read the screen as a list of elements and tap, type and install apps through MCP, a CLI, an SDK or an API.
How much does Phonebox cost?
$0.06 per phone-minute, billed per second, from plan credit. Starter is $20 a month and Pro is $99 a month, each including that much phone time. New accounts get a free week with $5 of phone time.
Does Phonebox work with Claude Code?
Yes. Phonebox has a hosted MCP server for Claude Code and other MCP clients, plus a CLI and an agent skill file. Your agent can set up the account itself from one prompt.
Does a Phonebox phone have a phone number?
No. Phonebox says it isn’t a phone number or an SMS inbox and doesn’t make or receive calls. It gives your agent control of the device itself.
Does Phonebox support iPhones?
No. Phonebox phones are virtual Android devices.
What happens when a phone is parked?
A parked phone stops billing and keeps its installed apps, files and signed-in accounts until it starts again.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
