Let Claude use an iPhone for you: a repo with 13 commits gained 700 stars in four days, and the stars came from the launch, not the code
What it is A remote MCP access layer. Paste a URL into Claude, ChatGPT, or Cursor, and the agent can operate a real iPhone or Android phone in the cloud. Behind it is a paid cloud-phone service.
Why it blew up The founder (ex-Browser Use growth lead) ran a demo of "Claude Code ordering a Waymo on my iPhone" and similar everyday scenarios, then launched the same product several times: a debut, a press round, an MCP version, and a YC Launch. 11-02 that post got 420K views, and 11-04 after the YC launch it gained +178 stars in a single day.
CHow you could follow it Copy the playbook for another niche: find a daily device or app that agents still can't reach, build a thin access layer that installs from a single prompt, add a remote MCP URL, write setup instructions per client, and use a 30-second everyday-life video to post across several X posts.
- Data (updated daily) · repo created 2026-10-23 · 744 stars
GitHub · stars Data 2026-11-04 → 2026-11-08 (5 days, 5 observations). Points before 10-28 come from deeply / archive backfill, not every day has one
phone-harness/mcp is the access layer from a YC Fall batch company: paste one URL into Claude, ChatGPT or Cursor and your agent can drive a real iPhone or Android phone in the cloud. The repo was created in Oct 23, has only 13 commits, and by Nov 6 already has 715 stars. The lesson for devs chasing stars: the same tech was open-sourced half a year earlier and only got 280 stars. The difference is all in how it was launched.
A video called Waymo
On Aug 8,21, 2026, 24-year-old Shawn Pana posted on X: "Claude Code can use your iPhone! 📲 Introducing: phone-harness." Three lines below: it operates any iOS app like a human, with no API and no jailbreak, connects once and is ready anytime, and "installs from one prompt." The attached video shows Claude Code calling a Waymo on his personal iPhone.
A commenter in the first batch of replies asked: "Is there an Android version? If not, I'd like to help build one." After that post, YouTube explainers, Instagram short videos and Claude user groups on Facebook kept reposting it. By Sep 2, the main repo phone-harness had 2,134 stars; by Nov 6 it was 4,573 stars (GitHub, 11-06). Notably, its Show HN on Aug 18 only got 2 points, so the traffic never came from developer communities.
What this repo actually is
It's thin. Written in Python, the files under src handle auth, cloud calls, the MCP server and Vercel deployment. Their job is to wrap Phone Harness's cloud phone API into a remote MCP endpoint, mcp.phone-harness.com/mcp. Users paste the URL, sign in via OAuth, and the agent can see the screen, tap, type and swipe. Self-hosting is also supported: local, Docker, or Vercel.
The bulk of the README isn't about how it works, it's setup instructions for 17 clients: Claude, ChatGPT, Cursor, Codex, Grok, Claude Code, VS Code, Windsurf, Devin, Gemini CLI, Copilot, Perplexity, Warp, Raycast, Manus, OpenClaw, Zed, plus one-click "Add to Claude / Cursor / VS Code" buttons. Users of each client can jump straight to their own section.
The iPhone path in the main repo is simple too: it uses Apple's built-in iPhone Mirroring, turning the phone into a window on the Mac. It reads the screen with system OCR and sends mouse and keyboard events to tap. No jailbreak, no Xcode, nothing to install on the phone.
How the stars came in: one product, launched four times
| Date | What happened | Volume |
|---|---|---|
| 08-08 | X launch, Claude Code calls a Waymo from a personal iPhone | Main repo reaches 2,134 stars by 09-02 |
| 10-15 | alphasignal writes it up ("over 3,600 stars") | Main repo +240 that day, +205 the next |
| 11-02 | X posts the MCP version: "Paste a URL, log in, and Claude, ChatGPT, Cursor and Codex can use it" | 420k views, 6.7K likes, 4.4K bookmarks (X, as of 11-06) |
| 11-04 | X posts "Agents can now control phones. Introducing Phone Harness (YC F26)" + YC Launch page | mcp repo 11-05 single-day +178, the highest day so far |
Under the 11-04 post sit three demos: Claude Code calling a Waymo on a personal iPhone, a cloud Android adding toilet paper to an Amazon cart, and Grok checking LinkedIn notifications on a cloud iPhone. None of them is about "testing apps"; they're all everyday phone tasks.
The mcp repo's own stars came almost entirely from the 11-02 and 11-04 launches: 11-04 464 stars, 11-05 642,11-06 715 (GitHub daily snapshots and API). It has no releases, and its issues are 0. Most of the stars are likely from followers of the main repo clicking through, which doesn't mean many people are actually using it.
Shawn previously led growth at Browser Use, managing product launches and developer partnerships. His earlier project smux (1,532 stars) followed the same playbook. "The phone version of browser-use" is an easy pitch to repeat, and YC gave it a second burst of exposure.
Where the money is: stars are the top of the funnel, cloud phones are the revenue
The open-source part is MIT and free. The paid part is the cloud phones on phone-harness.com: Android emulators at 5 US cents per minute, with $5 free for new accounts; Plus at $20/month includes 1,000 minutes; cloud iPhone at $80/month ($64 billed annually); a version with an eSIM real phone number at $100/month; and dedicated real devices above that. The pricing page sells "real hardware, apps think a real person is using them", "real phone number, handles SMS and two-factor auth", "login state kept persistently", and "one device to a hundred". The top tiers also mention "one phone, multiple identities" and "configurable proxies", which suggests they're aimed at running multiple accounts and receiving verification codes (my inference).
Revenue isn't public, and how many people pay is unknown.
Who it's competing with
| Project | Starting point | Scale | How it differs |
|---|---|---|---|
| mobile-next/mobile-mcp | 2025-03 | 10,317 stars, +25~35 per day | The mobile MCPs that launched first focus on app testing |
| droidrun/mobilerun | 2025-04 | 9,971 stars | Mainly Android, with a built-in agent |
| zai-org/Open-AutoGLM | 2025-12 | 26,546 stars | Zhipu's self-trained phone agent model, Android |
| minitap-ai/mobile-use | 2025-08 | 3,364 stars | Built-in agent, testing-oriented, growth has flattened (+3~6 a day) |
| CursorTouch/Android-MCP, appium-mcp | 2025-06 | 917 / 518 stars | Android or Appium only, you set up the environment yourself |
| jfarcand/mirroir-mcp | 2026-02 | 280 stars | Same trick (iPhone Mirroring driving a real iPhone), but about half a year earlier |
(GitHub data, star counts as of 11-06.)
The category is neither new nor empty: the top two each have more than twice the stars of Phone Harness's main repo. The most telling row is the last one. mirroir-mcp did the same thing with iPhone Mirroring back in February, with no demo videos, no cloud, and no launch, so it stalled at 280 stars. The technology isn't the moat. Phone Harness won on three things: its demos showed everyday life rather than testing; it sold "give your existing agent a phone" instead of building yet another agent; and a founder who knows how to launch shipped the same product four times at the right moment with YC.
Copycats exist but are small: DeepSeek-phone-harness (15 stars), phone-harness-safe (5 stars), and a repo that uses iPhone Mirroring to auto-post to social media (28 stars).
If it were you
Verdict: C. You can learn a repeatable playbook for earning stars, but building another open-source "agent controls a phone" project from scratch isn't worth it. Three established projects already sit at the 10k-star level, and this one relied on a founder's X following, Browser Use experience and YC, none of which you can get.
Who it suits: indie developers who already have some following in an AI client community and can cut a 30-second demo video; anyone looking for "the gap agents can't reach yet" to build a thin access layer for.
Who it doesn't suit: anyone who wants to compete head-on for stars in phone automation (mobile-mcp, droidrun and Open-AutoGLM are already there); anyone who only writes code and won't make their own posts and demos (mirroir-mcp is the proof); anyone building around "real numbers, multiple identities, apps think it's a real person", since platform bans and compliance risk haven't been tested.
How to do it: copy the playbook, change the target. Pick something agents can't touch yet but people use every day (smart TVs, car infotainment, a service that has an app but no API, old desktop software on Windows), build a thin access layer that installs from a single prompt, and give it a remote MCP URL. Write the README as one setup section per client plus one-click buttons. If you do go into phones, go where it hasn't been done: local Android real-device runs, or ready-made flows for non-US apps (China's super-apps, Grab and Shopee in Southeast Asia).
How to get users: a video under 30 seconds on X, with your own device doing something everyone understands (calling a ride, ordering food, grabbing a ticket), opening with "Claude can now use your XX". Split one project into several launches: the first release, a repost when someone writes it up, and another when the MCP version ships. Don't count on HN or Product Hunt; neither worked this time.
When to move: while "paste a remote MCP URL and it works" is still new. Ship the first version within three months. Once every AI client has phone and computer control built in, the novelty of this kind of access layer is gone.
What would show you're wrong: (1) phone-harness/mcp's daily star gains drop to single digits a month later and its issues are still 0, meaning the stars came from the launch and the playbook doesn't retain users; (2) Apple restricts automation of iPhone Mirroring, or Claude or ChatGPT officially ship phone control, which makes thin access layers pointless; (3) the reverse: similar projects without marketing, like mirroir-mcp, start growing without any launch, meaning the demand is real and building directly on phones is worth more than this assessment suggests.
Research notes: competitor table, acquisition evidence, full research report
- Competitors
Name Where Launched Scale How it differs from this one mobile-next/mobile-mcp + Mobile Next Cloud Open source 2025-03 10,317 stars, +25~35 per day A mobile MCP that launched first, leaning toward app testing. Phone Harness is about letting an agent use a phone as a user would, and it also sells cloud iPhones with real phone numbers. droidrun/mobilerun Open source 2025-04 9,971 star Mainly Android, with its own agent framework. Phone Harness ships no agent, only hands and eyes, and connects to the user's own Claude or Codex. minitap-ai/mobile-use Open source 2025-08 3,364 stars, growth has flattened Built-in agent, testing-oriented zai-org/Open-AutoGLM Open source 2025-12 26,546 star Self-trained phone agent model, Android. Not about plugging a phone into an existing agent. jfarcand/mirroir-mcp Open source 2026-02 280 star Drives a real iPhone via iPhone Mirroring too, and earlier. No marketing and no cloud, so it never took off. CursorTouch/Android-MCP / appium-mcp Open source 2025-06 917 / 518 star Android or Appium testing stacks only, you set up the environment yourself callstack agent-device / AgentCloud / Limrun / Kobiton Other 2026 Not verified Test clouds and emulators for mobile developers, not agent-oriented phones with numbers Gemini's on-device Android agent / Siri App Intents Big-company features — — Limited to their own ecosystems and requires app integration. No option for any agent to remotely open an iPhone. - Where users come from
- Verified Founder's X launch posts + demo videos: 8-08 main repo launch post (Claude Code calls a Waymo on an iPhone); 11-02 MCP launch post, 420k views / 6.7K likes; 11-04 YC launch post, mcp repo +178 the next day
- Verified YC backing / Launch YC: ycombinator.com/launches/WdR-phone-harness… (11-04). YC's company listing calls it the most popular open-source phone automation harness.
- Verified Tech media and secondary spread: alphasignal 10-15 write-up; main repo +240 that day and +205 the next; YouTube explainers, Instagram, Facebook Claude groups, Medium, thakicloud, kocpc
- Verified Agent ecosystem listings: listed on the Nous Hermes Agent Marketplace; the README gives setup instructions and one-click add buttons for 17 MCP clients
- Inferred HN / Product Hunt: Show HN 8-18 got only 2 points; no PH launch found (our hackernews / producthunt databases)
phone-harness/mcp: why it took off
Research date 2026-11-06. Data card: github repo, created 2026-10-23, 695 stars / 51 forks at time of retrieval, 11-05 single-day +178.
One line
phone-harness/mcp is the remote MCP access layer for Phone Harness (YC F26, solo founder Shawn Pana). Any MCP client, including Claude, ChatGPT, Cursor, Codex and Grok, can paste a URL (https://mcp.phone-harness.com/mcp), sign in via OAuth, and have the agent operate a real cloud iPhone or Android emulator on Phone Harness Cloud. It isn't a standalone product; it's the commercial entry point for the main repo phone-harness/phone-harness (4,570 stars). The stars came from the founder's two X launch posts and the YC Launch, not from this repo growing on its own.
1. What it actually is
- The code is thin: 13 commits, all written by ShawnPana alone.
src/phone_harness_mcp/contains auth, cloud, server and vercel modules. Essentially it wraps the Phone Harness Cloud API as a Streamable HTTP MCP server, and also supports local stdio, Docker and self-hosted Vercel deployment (GitHub API, 2026-11-06). - The README is mainly setup instructions for 17 clients (Claude, Cursor, ChatGPT, Grok, Claude Code, Codex, VS Code, Windsurf, Devin, Gemini CLI, Copilot, Perplexity, Warp, Raycast, Manus, OpenClaw, Zed), plus one-click "Add to Claude / Cursor / VS Code" buttons (github.com/phone-harness/mcp).
- Paid usage is in the cloud (pricing at phone-harness.com): Android emulator ¢5/min, new accounts get $5 free; Plus is $20/mo with 1,000 min included; cloud iPhone $80/mo (or $64/yr billed annually), with an eSIM real phone number $100/mo; the Scale tier is a dedicated real device. The pitch: "Stealth: real hardware, apps think it's a person", "Real phone number, SMS and 2FA handled", "Login state persists across sessions", "One to a hundred phones".
2. Competitors
| Name | Form | Launched | Scale | How it differs |
|---|---|---|---|---|
| mobile-next/mobile-mcp + Mobile Next Cloud | Open-source MCP + cloud real devices | 2025-03 | 10,317 stars (our GitHub library 11-06), gaining 25–35 per day | The first mobile MCP leader, positioned for testing/development (emulators, real devices, app testing). Phone Harness is positioned as letting an agent use a phone like a user (booking rides, shopping, receiving 2FA), and also sells an "agent's own phone" with a number. |
| droidrun/mobilerun | Open-source agent framework + cloud | 2025-04 | 9,971 star | Android-first, ships its own agent. Phone Harness ships no agent; it only provides hands and eyes, and your own Claude/Codex drives it. |
| minitap-ai/mobile-use | Open source | 2025-08 | 3,364 star | Ships its own agent, testing-oriented; growth has flattened (3–6 per day) |
| zai-org/Open-AutoGLM | Open-source model + framework | 2025-12 | 26,546 star | A Chinese phone-agent model, Android-focused; it trains its own model rather than "plugging into your agent". |
| jfarcand/mirroir-mcp | Open source MCP servers | 2026-02 | 280 star | The same trick (macOS iPhone Mirroring to drive a real iPhone) came earlier; without marketing or cloud, it never took off, which shows the technique itself isn't the moat. |
| CursorTouch/Android-MCP, appium/appium-mcp | Open source MCP servers | 2025-06 | 917 / 518 star | Only covers Android or the Appium testing stack; you have to configure adb/Appium yourself. |
| callstack agent-device, AgentCloud, Limrun, Kobiton, BrowserStack, etc. | Open-source CLIs / cloud emulators / testing clouds | 2026 years | Not verified | All run test suites for mobile developers; none sell an "agent phone with a real number". |
| Big-company features | Gemini/Android on-device agents, Apple Siri App Intents | — | — | All stay inside their own ecosystems and require app integration; there's no official way for any agent to remotely drive an iPhone. |
Bottom line: the category isn't new or empty, and mobile-mcp and droidrun each have more than double the stars. Phone Harness did not get "first".
3. How the growth happened (with evidence)
Timeline (star data from our GitHub library repo_day and the GitHub API):
- 2026-08-08 main repo launch: Shawn posted on X, "Claude Code can use your iPhone! 📲 Introducing: phone-harness", demoing Claude Code calling a Waymo on his own iPhone (x.com/shawn_pana/status/2085953331751776745). Over the following 8/8–8/10 days came YouTube explainers, Instagram, and reposts in Facebook Claude communities; thakicloud (8/10) wrote a hands-on review. By 09-02 the main repo had 2,134 stars. The Show HN (8/18) got only 2 points, so the traffic did not come from HN (our hackernews library). No Product Hunt launch found (producthunt library).
- 10-15/16 main repo jump (+240, +205, repo_day): the same day alphasignal.ai ran a story ("passed 3,600 GitHub stars").
- 11-02 MCP version launch: X post "Introducing: Phone Harness MCP 🔌…works in Claude, ChatGPT, Cursor, Codex…nothing to install, just paste a URL", with 420K views, 6.7K likes, 4.4K bookmarks, 418 reposts (x.com/shawn_pana/status/2117308255517249597). The mcp library first shows up in our data that day, with 11-04 and 464 stars.
- 11-04 YC Launch: X post "Agents can now control phones. Introducing Phone Harness (YC F26)" plus a YC Launch page (x.com/shawn_pana/status/2118010388621000770; ycombinator.com/launches/WdR-…). The next day the mcp library gained 178 (11-05, the peak day on our data card). Taiwan's kocpc and other outlets followed with write-ups.
- The main repo moved from ShawnPana/phone-harness to the phone-harness organization. Our library kept the old name until 11-02, which corresponds to 4,051 stars; the GitHub API now shows 4,570.
Channel attribution:
- Founder's own X account + demo video: the main channel, confirmed. Shawn was previously Head of Growth at Browser Use, handling product launches and developer partnerships (ycombinator.com/founders/2048642-shawn-pana); his GitHub bio says he is 21 years old, does open source, and worked on browser-use. His earlier smux (1,532 stars) and aurl (169) used the same playbook. This is a person who knows how to launch doing the launching.
- YC backing / Launch YC: confirmed (F26 batch). YC's AI company listing describes it as "the most popular open-source harness for phone automation".
- Secondary spread: YouTube explainers, an Instagram reel, Facebook groups, Medium, alphasignal, thakicloud, zentor, hermes-ai.net, and kocpc all confirmed, but each came after the X post.
- Ecosystem placement: listed in Nous's Hermes Agent Marketplace (hermes-agent.nousresearch.com/marketplace/phone-harness), and the README covers 17 clients, which makes it easy for each community to share. Confirmed.
- Ads: none found (our ad libraries don't track this kind of developer tool, so not finding ads doesn't mean none were run).
4. Why it took off
- The demos are well chosen and shareable: Claude Code calling a Waymo on a personal iPhone, adding toilet paper to an Amazon cart, checking LinkedIn notifications. "An agent using a phone for me" is instantly understood. Competitors like mobile-mcp pitch testing.
- Riding Apple's existing channel to bring the barrier to zero: iPhone Mirroring turns the phone into a window on the Mac; it reads the screen with Vision OCR and clicks by sending HID events. No jailbreak, no Xcode/WebDriverAgent, nothing installed on the phone (alphasignal 10-15). Real iOS device automation used to be the hardest part; now a single setup prompt gets it running. mirroir-mcp used the same approach back in February, so the technology isn't the reason. Packaging and launch are.
- Timing: by 2026, MCP had become the standard interface supported by Claude, ChatGPT, Cursor, Codex, and Grok. Remote connectors can take a URL directly and use OAuth. The mcp library is the "paste a URL and go" step, arriving just as every major platform opened custom connectors.
- Open-source for traffic, cloud for revenue: the open-source main repo (MIT, 4.5K stars) earns attention and trust; the mcp library funnels that traffic straight into paid cloud phones (per-minute billing / iPhone subscriptions). Stars are the top of the funnel; the money is in Cloud.
- Founder background + YC: a former Browser Use growth lead, so the "Browser Use for phones" story sells itself; YC gave a second burst of exposure.
5. Differentiation
- Versus mobile-mcp / droidrun / mobile-use: those are "test your app for developers" or "agent framework with its own agent". Phone Harness is "give your existing agent a phone": no agent included, it only sells the phone's hands and eyes, plus a real number (eSIM), the ability to receive SMS/2FA, long-lived login state, and real hardware that "Apps see a person". This pitch is clearly aimed at phone farms, multi-account operations, and 2FA-gated workflows (the pricing page's Scale tier lists "Multiple profiles per phone" and "Proxies of your choice").
- Versus mirroir-mcp: the technology is nearly identical. The difference is launch capability, cloud, and multi-client setup docs.
- Is it all launch execution? Partly. The "agent as user on a real iPhone" story determines whether the content spreads, and that is a real differentiator. But the core technique (iPhone Mirroring + OCR) had prior art, and what actually pushed it out was the founder's X launch skills and YC timing. The mcp library's stars came almost entirely from the 11-02 and 11-04 launches. Whether they stick depends on whether anyone pays for the cloud phones.
6. How many people have seen it
- Revenue not public (no TrustMRR listing found, no founder revenue posts found).
- Widely seen already: the MCP launch post got 420K views, there's a YC Launch, and alphasignal, kocpc, Medium, and YouTube have all covered it. There's pushback too, for example @sxtvik complaining "why is yc funding all these stupid companies why do you need a phone…" (x.com/sxtvik/status/2118051214017380672). Our x-radar corpus doesn't include his post, which suggests it hasn't reached the big accounts we track.
- Copycats: GitHub already has forks and lookalikes reusing the name, such as 2903077918-lgtm/DeepSeek-phone-harness (15 stars, 8-15), coopersimson96/phone-harness-safe (5 stars, 8-11), and ZachDreamZ/universal-phone-harness; there's also resocket-iphone-mirroring-social-media-publish (11-01,28 stars), which uses iPhone Mirroring for social-media publishing automation. All are small.
7. What to borrow
- Open-source a very thin harness that installs from a single prompt to gather attention, then use a "remote MCP URL + OAuth" to send open-source users straight to a usage-billed cloud service.
- The README is structured as "setup guides for 17 clients + one-click buttons", so each client community can share it on its own.
- The demos feature everyday life (calling a ride, shopping, 2FA), not developer testing scenarios, which widens the sharing radius by an order of magnitude.
- One product, several launches: open-source debut, press coverage, MCP version, YC Launch, each time re-promoted on X with video.
8. Still unclear
- How many paying cloud-phone users there are and how much revenue they bring: unknown.
- Is the mcp repo's star count real usage? The repo has no releases and no issues, only 13 commits. Most of the stars were probably tossed in by fans of the main repo. The GitHub API returned 404 for the per-day stargazer breakdown, so I couldn't get it.
- Will Apple restrict iPhone Mirroring from being used this way (it's already unavailable in the EU)? That's the platform risk for this iOS route.
- Will selling points like "stealth / real phone number / multiple profiles" trigger anti-automation bans or platform compliance issues in individual apps? Unverified.