Reval Labs · Opposable: Computer Use for Phones, Letting Claude Code, Codex, and MCP Agents Tap and Type in Android and iPhone Apps
Opposable is computer use for phones: an Android app (plus an iPhone beta driven from a Mac) that exposes the screen and taps, swipes, and typing as MCP tools, so Claude Code, Codex, Cursor, any MCP client, or the Claude and ChatGPT apps through a connector can work inside the apps you are signed into. It suits people who already use AI agents and want them to finish jobs that dead-end in a phone app, such as a 2FA code, a banking task, or testing their own app on a real device.
Best for
Developers and power users who already run Claude Code, Codex, or another MCP agent and want it to operate real phone apps (2FA codes, banking with approval, messaging, rides) or test their own Android app on a physical device
Not ideal for
People who want a polished consumer assistant installed from an app store, teams that need audited security controls before letting software operate signed-in banking or work apps, iPhone-only users without an Apple silicon Mac or willingness to flash a stick, and anyone who cannot leave a phone unlocked during runs
Who it's for
Developers and technical users of coding agents such as Claude Code and Codex who want their agent to operate Android or iPhone apps, including mobile developers testing on real devices
Opposable is a practical answer to a real gap: coding agents can drive browsers and desktops, but the jobs that stall them, such as a 2FA code, a banking app, or a ride booking, live on the phone. Its design choices are sensible for that job. It gives the agent a labelled element list instead of only pixels, holds an outbound connection instead of opening ports, runs common tasks with no model, and requires an Approve tap for payments. The free tier is genuinely useful if your agent and phone share a network, and it now includes scheduled tasks and the iPhone beta; the Claude and ChatGPT connector widens the audience beyond coding-agent users, though after a 14-day trial it needs Pro. The costs are trust and maturity. You sideload an accessibility app, keep the phone unlocked, guard a token that is effectively a password to your phone, and route screen contents through a relay if you want remote access, all from a team that launched in October 2026 and has not published an API reference or certifications. Its own report also shows phones dropping off the relay when Android's battery managers close the app. Start with a spare phone or a cloud phone and low-stakes apps, and keep payment approvals on before you point it at accounts that matter.
Who should use it
Developers and technical power users who already use Claude Code, Codex, or Cursor and keep hitting tasks that end in a phone app; mobile developers who want their coding agent to test builds on a real Android device; and people willing to dedicate a spare Android phone or rent a cloud phone as an agent's phone.
Who should skip it
Non-technical users who expect an app-store install and a finished assistant, organizations that need documented security reviews before software operates signed-in work or banking apps, iPhone users without an Apple silicon Mac who do not want to flash hardware, and anyone uncomfortable granting an app full accessibility control of their main phone.
Free
$0
Pro
$8
Billed monthly
Power
$29
Billed monthly
Cloud Starter
$19
Billed monthly
Cloud Always-on
$49
Billed monthly
Cloud Team
$149
Billed monthly
Free tier limits: The free plan covers everything that runs on your phone, including scheduled tasks and payment approvals, with unlimited tasks over USB or the same Wi-Fi. Each phone gets the relay free for 14 days; after that, reaching the phone from anywhere (including from the Claude and ChatGPT connectors), webhooks from your own server, and recipe sync need Pro or Power.
Note: Checked on the Opposable website on November 10, 2026. Pro is marked as a founding offer. Some Pro and Power features are listed as coming: automatic recipe repair, priority recipe repair, and a shared team recipe library. Early cloud-phone subscribers keep their launch price while they stay subscribed, and the pricing page offers a refund if a cloud phone does not work for you in the first 3 days. The terms (last updated October 8, 2026) say subscriptions renew monthly or yearly, the current period is not otherwise refunded on cancellation unless the law requires it, and cloud phones are deleted within 30 days after a subscription ends. The optional iPhone stick for Windows or Linux is a third-party ESP32-S3 board that costs about $8 to $13.
API pricing
Free local MCP server and REST API; remote access through the hosted relay, including the Claude and ChatGPT connectors, is free for 14 days per phone and then requires Pro ($8/month) or Power ($29/month)
Available models
Getting past a verification code mid-task
When a sign-in sends a one-time code by text, the phone reads it from Messages or the notification and types it where it is needed, so the agent can finish instead of stopping to ask you.
Testing your own Android app on a physical device
A coding agent connected over MCP can open the app it just built, tap through a flow, and read the labelled screen elements, on a real phone rather than a simulator.
Delegating a recurring phone chore
A scheduled task, which the phone runs by itself on the free plan, wakes the phone, runs a saved recipe with no model, and messages you the outcome, which suits checks like parcel status that only exist in an app.
Opposable vs. ARTEMIS
ARTEMIS is Google's open-source (Apache 2.0) Android test automation agent: it runs on your computer, drives phones or emulators over ADB, uses model keys you supply (Gemini by default), and targets QA and bug reproduction. Opposable runs as an app on the phone, reaches it over USB, Wi-Fi, or a hosted relay, includes a free on-device model, and targets everyday tasks in signed-in apps with payment approvals, plus an iPhone beta and cloud phones.
Opposable vs. Cua
Cua gives agents desktops: an MIT-licensed driver for native macOS, Windows, and Linux apps, plus local and cloud sandboxes (including local Android emulators) for evals and training. Opposable focuses on phones, especially your own physical Android phone or iPhone with your real signed-in apps, and adds phone-specific pieces such as reading one-time codes and approving payments on the device.
Opposable vs. Browser Use
Browser Use automates websites through an open-source browser agent library and cloud browsers. Opposable operates native phone apps, many of which have no web or API equivalent. A team could use Browser Use for web flows and Opposable for steps that only exist in a mobile app.
What is Opposable?
Opposable is computer use for phones, made by Reval Labs (run by 100x Ventures OÜ in Estonia, according to its privacy policy). An Android app (and an iPhone beta driven from a Mac) lets Claude Code, Codex, Cursor, any MCP client, or the Claude and ChatGPT apps read the phone's screen and tap, swipe, and type in the apps you are signed into. It also has a free on-device agent, approval taps for payments, and rentable cloud Android phones.
How much does Opposable cost?
The phone app, its MCP server, the REST API, scheduled tasks, payment approvals, and the iPhone beta are free over USB or the same Wi-Fi. Every phone gets the hosted relay free for 14 days with no card. After that, Pro is $8 a month or $79 a year (founding members get $5 a month for life) and covers the relay for Claude Code and Codex, the Claude and ChatGPT connectors, webhooks from your own server, and recipe sync across up to 3 phones. Power is $29 a month for up to 10 phones, with an audit log across them. Cloud phones are $19, $49, or $149 a month.
Does Opposable work on iPhone?
Yes, in beta. Opposable for Mac pairs with the iPhone as a Bluetooth keyboard and pointer and sees its screen over a USB-C cable or AirPlay, so nothing is installed on the phone and no jailbreak is needed. It needs an Apple silicon Mac on macOS 14 or later, an iPhone on iOS 18 or later, and AssistiveTouch turned on. The Mac page says the phone's screen goes from the cable or AirPlay to your AI app without passing through Opposable's servers. On Windows or Linux, an ESP32-S3 stick you buy for about $8 to $13 and flash with Opposable's beta firmware does the Bluetooth part. Both the Mac app (0.2.1-beta) and the stick firmware (0.2.0) are betas, the Mac app is not yet notarized by Apple, and Opposable says it is still testing the firmware across iPhone models. The pricing table lists iPhone through a Mac on the free plan.
Which agents can control the phone?
Anything that speaks MCP. Claude Code connects with one `claude mcp add --transport http` command, Codex takes a few lines in ~/.codex/config.toml, and the pairing page and the `opposable connect` command-line tool also give the setup for Cursor and other clients. The Claude app (web, iPhone, and Android) and ChatGPT in Developer mode can add an Android phone as a custom connector at getopposable.com/mcp. The Mac app can add itself to Claude Code or Claude Desktop. You can also hand tasks to the phone through a REST endpoint and webhooks.
Do I need a model API key?
No. Claude Code or Codex do the reasoning on the subscription you already have, and the agent built into the app runs on the phone for free with its own on-device model. The app can optionally use your own API key or a model on your network.
Is it safe to let an agent operate my phone and banking apps?
Opposable documents several safeguards: anything that looks like a payment or purchase waits for your Approve tap on every plan, actions go into an audit log, and the server listens only on localhost unless you allow the network. Whoever holds the token can operate the phone, so treat it like a password. With remote access on, requests and screen contents pass through the relay in transit, though the privacy policy says they are not stored. The terms note that agents make mistakes and ask you to keep approvals on.
What is an Opposable cloud phone?
A private virtual Android 15 phone (2 vCPU, 3 GB RAM) that you sign into through a browser viewer and connect to your agent over an HTTPS MCP endpoint. Cloud Starter is $19 a month with 60 active phone-hours, then $0.15 an hour, and sleeps when idle; Cloud Always-on is $49 a month; Cloud Team is $149 a month for three always-on phones. Some banking apps refuse to run on virtual phones.
Opposable vs ARTEMIS: what is the difference?
Both let coding agents operate a real Android phone over MCP. Google's ARTEMIS is an open-source test automation agent that drives devices over ADB from your computer with your own model keys, aimed at QA and bug reproduction. Opposable is an app on the phone itself, aimed at doing everyday tasks in your signed-in apps, with payment approvals, an on-device model, a hosted relay, an iPhone beta, and rentable cloud phones.
Android developers and mobile QA engineers who want an AI agent, driven from their coding assistant or test suite, to operate real devices for testing and bug reproduction
Free
Wang Zan (wangzan101)
Technically comfortable Android 14+ users who want a phone agent that operates their own signed-in apps and websites, prefer an open-source app with no vendor server, and are willing to sideload an APK, enable developer settings and supply their own model
FreeMost computer-use tools stop at the browser or the desktop, while banking, rides, messaging, and the one-time code in the middle of every sign-in live on the phone. Opposable fills that gap. On Android 11 or newer you sideload a 43 MB APK, switch it on as an accessibility service, and pair it with your computer; one `claude mcp add --transport http` command (or a few lines in Codex's config.toml, or the `opposable connect` command-line tool) turns the phone into a tool in your next agent session. The Android phone can also be added to the Claude or ChatGPT app as a custom connector, so you can ask from a chat on the web or on the phone itself. The agent gets each screen as a list of labelled elements with positions plus a screenshot when it needs one, and calls tools such as tap_element, swipe, type_text, launch_app, and press_key, each of which reports what it actually did. The phone keeps one outbound connection, so it works over USB, the same Wi-Fi, or from anywhere through Opposable's hosted relay without port forwarding; the relay, which the Claude and ChatGPT connectors always use, is free for 14 days per phone and then needs Pro. There is also a delegation mode: a REST call or MCP request hands the whole task to the phone, which runs it locally and returns done or not done, can run on a schedule that the phone keeps by itself, and saves repeated tasks as recipes that replay without a model. About 40 everyday tasks (alarms, calendar, texts, settings) run with no model at all, and an on-device screen model (a 2-billion-parameter GUI-Owl model run through llama.cpp) handles the rest offline, so you can use it without any API key. Anything that looks like a payment pauses for an Approve tap on the phone, and actions are written to an audit log. iPhone support is a beta that needs no app on the phone: the Opposable Mac app (Apple silicon, macOS 14+) pairs with the iPhone as a Bluetooth keyboard and pointer and sees the screen over a cable or AirPlay, and on Windows or Linux a small ESP32-S3 stick flashed with Opposable's firmware does the Bluetooth part. If you have no spare phone, Opposable rents private virtual Android 15 phones from $19 a month. The phone app, its MCP server, the REST API, scheduled tasks, payment approvals, and the iPhone beta are free; Pro ($8/month) and Power ($29/month) pay for the relay after the trial and for recipe sync across phones. The tradeoffs are real: the Android app is not on the Play Store, unattended runs need the phone charging and unlocked, the bearer token is effectively a password to your phone, relay traffic (which can include screen contents) passes through Opposable's servers, phones can drop off the relay when Android's battery management closes the app, some banking apps refuse to run on the cloud phones, and the product launched on October 7, 2026, with the iPhone path explicitly in beta.
Opposable published usage data from October 8 to November 3: 91% of agent actions worked and 99% of taps and swipes did what was asked, but 27 of 45 phones dropped off the relay at least once, with Xiaomi and Samsung phones dropping off most often. The app now runs as a foreground service, asks to be left out of battery optimisation, and reconnects on a 15-minute alarm.
A technical post described the built-in agent as GUI-Owl 1.5 2B quantised to 4 bits on llama.cpp, with a step taking 7.3 seconds on a laptop CPU and the vision encoder running in about 0.24 seconds on a Galaxy S25's NPU.
Within days of launch, Opposable added a custom connector so the Claude and ChatGPT apps can use an Android phone, gave every phone a 14-day free trial of the relay, and moved scheduled tasks and iPhone through a Mac to the free plan; the pricing page now says payment approvals are free as well. Pro now pays mainly for remote access and recipe sync.
Reval Labs introduced Opposable in its "Giving phones thumbs" launch post: a free Android app that exposes the phone's screen and touch input as MCP tools for Claude Code, Codex, Cursor, and other agents, with payment approvals, an on-device model, and an iPhone version in beta. A Show HN post followed on October 8.
Are you the founder? Claim this listing →