mrkeyoor.comAsk

JosssphZhou/pokitested 19 Nov 2026commit d5711d2

poki review

Use Poki for experiments with personal iPhone tasks if you can manage signing, pairing and checks in the target app.

  • Safety failed
  • Tests failed
4 sto install
  • 35 packages
Automationevaluationupdated 21 Nov 2026
trackingstars / 7d

Poki is an open-source iPhone agent that reads the screen and taps, types and scrolls inside apps using your model account. Its README documents installation from a Mac or Windows PC, standalone operation afterward, and confirmation for payments, deletions and Wallet actions.

Verdict

Use Poki for experiments with personal iPhone tasks if you can manage signing, pairing and checks in the target app. Avoid it for time-sensitive unattended errands: issue #165 records background hangs, and issue #124 records restart failures on Poki 0.2.9 (672) with iOS 18.7.7. Our source run installed dependencies in 4 seconds and failed its tests after 399 seconds; the later deployment attempt stopped at its safety check.

Poki infographic: operates iPhone apps using your model account, plus its uses, our test result, pitfalls and who it suits
Tested 19 November 2026: the source install took 4 s.

We ran it

Lab card: what happened when we ran poki
Lab card from our sandbox run
Screenshot of poki (poki.bot)
poki.bot
Install✓ · 4s35 packages · 47 MB
Buildn/ano build script
Tests✗ · 399sran, no count parsed
Known vulns10 critical · 1 high · 0 moderate · 0 low (npm audit)
Repo1677 files~150,029 lines of source · 46.7 MB · 21 CI workflows · tests dir

Answers from our run

Does poki build from source?

Dependencies installed in 4 seconds (35 packages), and the project has no separate build step. We cloned commit d5711d2 into a clean Debian container with 3 CPUs and no project-specific setup.

Do poki's tests pass?

No. The project's own test command exited unsuccessfully after 399 seconds, and its output did not expose an individual test count.

Does poki have known vulnerabilities in its dependencies?

npm audit flagged 1 known advisory in the dependency tree at the time of our run.

Who should not use poki?

People relying on time-sensitive unattended errands: issue #165 reports background hangs on Poki 0.3.0 (746) Alpha.

What are the alternatives to poki?

Appium, Maestro. Use Poki for experiments with personal iPhone tasks if you can manage signing, pairing and checks in the target app.

Setup4/5Our source checkout installed in 4 s with 35 dependencies, but we did not run a README task.Editor’s note: The npm dependency check succeeded, but installation on a supported iPhone was not reproduced.
Docs4/5The README has an install section (“Install Poki on your iPhone”) and a usage section (“How to use”); we did not run its quickstart ourselves.Editor’s note: The English README and verification guide were reviewed, but their complete native setup workflow was not executed.
Community4/5Last push on 2026-11-21, 1,706 stars, maintainers answered or closed 10 of the 10 most-discussed issues we sampled and 114 open issues and pull requests is a heavy backlog for its size.Editor’s note: Issue threads contain maintainer responses and reporter follow-ups, without a measure of adoption or support capacity.
Maturity1/5Latest release app-0.4.6-beta.1 on 2026-11-20 (pre-1.0), repository created 2026-11-05 and its test suite did not pass in our sandbox.Editor’s note: Source tests failed and deployment stopped at safety; production phone reliability was not measured.

How we score

Who it’s for

iPhone owners who want to experiment with delegated app tasks and can manage Apple signing, pairing and external connection settings.
Developers who want to inspect phone-control code and check results in the target app.
People prepared to compare issue reports with their installed build, operating-system version and device state.

Who it’s NOT for

People relying on time-sensitive unattended errands: issue #165 reports background hangs on Poki 0.3.0 (746) Alpha.
iOS 18.7.7 users who require dependable recovery after restarting: issue #124 reports a developer-image mount failure on Poki 0.2.9 (672), despite attempted USB re-pairing and other recovery steps.
People choosing Poki for continuous Screen History: issue #484 describes empty capture and recording that stops.
Anyone requiring all screen processing to stay on the iPhone: the verification guide says screen contents go to the configured model service.

Setup reality

Our November 19, 2026 run at commit d5711d2 installed npm dependencies in 4 seconds, recording 35 dependencies and 47 MB on disk. The build was skipped because there was no build script or target. Tests failed with exit 1 after 399 seconds, and npm audit reported one high-severity vulnerability. The test tail contains ERR_ASSERTION with actual 1 and expected 0.

Phone installation requires a separate workflow. The README lists an unlocked iPhone over USB, a Mac or Windows PC, an Apple Account, Developer Mode, Automation Mode, an external connection and a model account. Free signatures last seven days. Issue #35 says first-time setup requires Wi-Fi.

The sandbox configuration was 3 CPUs, 8 GB RAM and image lab-node:22. A November 20 deployment safety check failed because no recognized Compose file existed. The scope and unresolved diagnostic details are collected in What we have not tested yet.

Our November 19 npm run installed dependencies in 4 seconds, then the test command failed after 399 seconds. Those are our source-run measurements. Use Poki for experiments with personal iPhone tasks if you can manage signing, pairing and checks in the target app. Avoid it for time-sensitive unattended errands: issue #165 reports background hangs on Poki 0.3.0 (746) Alpha, including a task that stalled after opening WeChat.

What people use poki for

Users report calculator control over cellular, a completed WeChat task and installation through Poki Setup; the Screen History report describes empty capture. The Poki README describes an open-source iPhone agent that reads the screen and operates real apps using your model account. The distinction between an advertised capability and a reported outcome matters here: the issue threads identify the build, device state or installation route behind each result.

Person and use Reported outcome Evidence
Zeno6898, Windows control over cellular Candidate build 506 reopened automation and produced 7×9=63, checked with a separate screen read. Issue #35: configured phone with its developer image already mounted.
Zeno6898, operating WeChat A restarted session completed a task in 89 seconds with 11 tool calls, but background hangs recurred. Issue #165: success and subsequent failure in the same report.
ZhehanZhang, graphical Mac installation A notarized Poki Setup 0.2.7 build installed on a real iPhone. Issue #484: contributor testing of build 1752 while PR #509 awaited review.
Joseph, Screen History Poki said it had seen nothing after several minutes using other apps. Issue #484: short capture runs and stopping conditions.

The calculator report includes a separate screen read after the commands finished. That gives a reader an observable result to compare with the requested task. The README uses the same practical approach for its first task: create a Notes entry titled "Poki Demo" containing Milk, Eggs and Coffee, leave existing notes unchanged, then open Notes to inspect the result. That is a documented starting task rather than a result from our run.

What happened when we ran it

Our November 20 deployment safety check failed because no recognized Docker Compose file existed; our November 19 source run installed dependencies and failed its tests. The deployment safety result and source-run measurements describe separate checks. The source checkout was commit d5711d2, with 1,677 files, approximately 150,029 source lines and 46.7 MB checked out. Its configured sandbox had 3 CPUs, 8 GB RAM and image label lab-node:22.

Question Answer Verdict Why
Did npm dependencies install? Yes, in 4 seconds; 35 dependencies and 47 MB on disk. Passed The install exited 0.
Did the build run? No build script or target existed. Skipped The harness skipped that step.
Did tests pass? No; exit 1 after 399 seconds. Failed The tail shows ERR_ASSERTION, actual 1, expected 0, strictEqual.
What did npm audit report? One high-severity vulnerability; zero critical, moderate or low findings. Finding recorded The audit total was 1.
What repository signals were recorded? 21 CI workflow files, a tests directory and no Dockerfile. Inventory recorded These are repository observations.
Did the November 20 deployment pass safety? No compose.yaml, compose.yml, docker-compose.yaml or docker-compose.yml existed. Failed The safety rule required a Compose file.
Did teardown finish? The record says ok. Completed The teardown step reported success.

The test tail also includes a passing case about reporting that a phone never saved recordings, with a duration of 317.363789 ms, followed by a failing-tests heading. The assertion records ERR_ASSERTION, actual 1, expected 0, operator strictEqual and diff simple. Installation, build and test observations belong to this npm workflow; they should be read alongside the native installation requirements rather than substituted for them.

The November 20 run summary says deployment used the repository's own Docker Compose configuration with every port bound to 127.0.0.1. Its preceding safety step records the missing Compose file and a failure. The failed safety step takes precedence when describing the outcome. The later teardown record reports ok; cleanup success does not turn the earlier step into a deployment success.

Where people get stuck: signing, pairing and background execution

Issue #124 documents a restart failure on Poki 0.2.9 (672) with iOS 18.7.7, including attempted USB re-pairing. Other threads describe Apple sign-in, Windows trust, connection forwarding and background execution failures. Their remedies apply to different stages: installing the app, reaching the phone's pairing service, mounting its developer image or keeping a task running after another app opens. The table keeps those symptoms separate.

Symptom and version Documented behavior Response in the thread
Apple sign-in HTTP 503 on Poki 0.4.5 Setup sent macOS where Apple expected Mac OS X. Issue #542: fixed releases and a later Mac/iPhone confirmation.
Windows installation stops at provisioning profiles Apple Devices continued showing Trust. Issue #421: back up and clear stale pairing records, reconnect and trust again.
Pairing fails despite an enabled Developer Mode Shadowrocket accepted a connection but stopped forwarding it. Issue #69: reconnecting Shadowrocket restored the reporter's connection.
Restart failure on Poki 0.2.9 (672), iOS 18.7.7 Developer-image mounting failed; USB re-pairing and other recovery attempts failed. Issue #124: diagnostic work followed.
WeChat task hangs on Poki 0.3.0 (746) Alpha Returning to Poki exposed the timeout. Issue #165: rollback to the 0.2.9 code in 0.3.1.
Earlier macOS helper stops before readiness Rebuilding overwrote the running helper. Issue #6: PR #59 replaced the binary; the reporter confirmed recovery.

The Windows reporter closed Apple Devices and Poki Setup, backed up %ProgramData%\Apple\Lockdown, deleted the files inside that folder, then reconnected the cable and accepted Trust. Apple Devices stopped showing its Trust button, and installation completed. The confirming comment in #421 identifies stale Windows pairing records as the cause. Keeping the backup is part of the documented remedy, which changes saved pairing state on the computer.

In #124, the runner cycled through connecting, pairing, mounting_ddi and waiting_network, reporting "mounting the developer disk image failed: device socket io failed". Attempts included reconnecting Shadowrocket, trying LocalDevVPN and Clash Mi, unlocking manually, resetting Location & Privacy, USB re-pairing, reinstalling Poki and updating to 0.2.9 (672). The restart report also records a ready relay and a missing iOS passcode prompt, making this a different symptom from #69's recovered forwarding connection.

The background report includes a wait of about 6 minutes before the reporter returned to Poki and saw do did not finish within 173 s. A separate computer-triggered calculator attempt also timed out until Poki returned to the foreground. Maintainers discussed audio recovery checks before announcing the rollback. Issue #165 records both the stalled steps and the later release response.

USB setup comes before standalone use

The README requires an unlocked iPhone connected over USB to a Mac or Windows PC for installation. Its requirements list iOS 18 or later, macOS 14 or later, Node.js 24 or later for the Mac agent installer, and Apple Devices or iTunes on Windows. The README installation sequence then asks users to accept Trust, enable Developer Mode and Automation Mode, configure an external connection and add a model account.

A free Apple Account works, and free signatures last seven days. Poki Setup renews those signatures while open, with phone sign-in also documented. The external connection routes 10.7.0.1 back to the iPhone through a tool such as LocalDevVPN, Karing or Shadowrocket. Model options include ChatGPT Plus or Pro sign-in, supported provider API keys and an OpenAI-compatible server. Those requirements explain why standalone operation still involves signing, connection and account choices.

English documentation and observable task results

Poki has an English README, and its verification guide defines sent: true as a touch written to the connection. The guide tells readers to inspect screen_changed and the returned screen because the device sends no acknowledgement. That distinction gives a concrete way to assess an apparent success: compare the target app's visible state with the requested outcome rather than treating a sent command as the outcome itself.

The verification guide separates simulator checks, physical-device measurements and protocol tests. Its protocol suite uses isolated state, ephemeral ports and external-process fixtures; it does not inject input into a phone or simulator. The README also labels Poki MCP and Poki Agent as earlier computer-driven tools that are no longer updated. These labels help identify which component a test result or older issue describes.

Screen contents and support bundles need separate privacy decisions

Poki's verification guide says screen contents go to your configured model service. The README separately says payments, deletions and Wallet actions stop for confirmation. These statements describe different parts of the workflow: the guide identifies the screen-data destination, while the README describes action approvals. Running the app on the phone therefore includes a model-service decision as well as an installation decision.

ScottH1122 explains in issue #369 that attachment text follows the normal message path and is retained for conversation history and retry. Joseph announced JSON, Markdown and plain-text attachments in Poki 0.4.2 Beta with a 100 KB per-file limit. Separately, issue #421 records a support archive containing local identifiers despite its redaction claim, followed by guidance to keep raw bundles private and review them before sharing.

Poki vs Siri, Xcode installation, Appium and Maestro

Poki targets delegated personal-phone tasks, while Appium and Maestro describe developer testing workflows. Appium is an open-source automation framework built around WebDriver. Maestro's README excludes physical iOS devices. The Appium README describes platform drivers and test clients; the Maestro README describes YAML flows and an MCP interface. Installing Poki from Xcode changes the installation route while keeping the same assistant.

Option Documented distinction Choose it when
Poki versus Siri An openmys X post positions Poki around complex tasks across apps. The particular personal-phone task warrants managing Poki's setup. Openmys's positioning.
Poki Setup versus Xcode Joseph documents source installation with personal signing and USB; free accounts use PokiExternal. You want to manage the native project and signing. Source-install instructions.
Poki versus Appium Appium uses WebDriver, platform drivers and programming-language clients. You want to write automation tests. Appium README.
Poki versus Maestro Maestro uses YAML flows and supports iOS simulators; physical iOS devices are excluded. Your testing target is supported by Maestro. Maestro README.

Outdated claims, corrected

The opening Apple-outage diagnosis in #542 was superseded by a client OS-name correction in PR #549. Joseph initially said switching machines or reinstalling would not help, then identified macOS where Apple expected Mac OS X. The later comments in #542 announce Poki Setup 0.2.8 and Poki 0.4.6 Beta, followed by huwan's successful Mac installation and iPhone sign-in report.

The openmys X claim about Wi-Fi and cellular needs the setup distinction in #35: first-time setup requires Wi-Fi, and candidate build 506's result concerns an already configured device. The remaining work moved to PRs #84 and #91. The deployment summary also needs correction because its safety step failed before deployment. Finally, Poki.com describes browser games; that service is separate from JosssphZhou/poki and its iPhone automation workflow.

Questions people ask

First-time Poki setup requires Wi-Fi; issue #35's successful cellular reports concern configured phones. The cellular discussion distinguishes preparing the developer image, restarting the phone, reopening Poki and switching an existing session to cellular. Those are separate conditions, even when they appear in one thread. The questions below follow reported installation problems and feature requests rather than treating every closed issue as the same kind of resolution.

Can Poki start automation over cellular with Wi-Fi and USB disconnected?

Candidate build 506 reopened Poki and completed 7×9=63 with USB disconnected and Wi-Fi off, according to Zeno6898. The phone already had its developer image mounted, and a separate screen read checked the result. Issue #35 moved the remaining work to PRs #84 and #91 and documented Wi-Fi for first-time setup. The reported result belongs to that configured device and candidate build.

Why does Poki ask me to enable Developer Mode when it is already on?

Issue #69 describes a pairing connection problem that presented as a Developer Mode hint. Reconnecting Shadowrocket restored forwarding for the reporter. Joseph explains that accepting the connection to 10.7.0.1 and forwarding it to the pairing service were different steps. The follow-up in #69 says PR #70 changed the hint to point readers toward Loopback and Pairing.

Why can Xcode see my iPhone while Poki lists no devices?

The older computer-tool report in #6 involved a native helper stopping before readiness after a rebuild. PR #59 changed replacement of the helper binary and added exit diagnostics; the reporter confirmed recovery. Issue #6's resolution concerns that helper path. The README identifies Poki MCP and Poki Agent as earlier computer-driven tools, distinct from the standalone app's installation workflow.

How do I fix Apple's HTTP 503 error when installing Poki?

Joseph advises users on 0.4.5 to install through Poki Setup 0.2.8 on a Mac once, then return to in-app updates from 0.4.6 onward. Huwan reported that 0.4.5 (1800) still failed, while Mac signing and installation of 0.4.6 (1922), followed by iPhone sign-in, succeeded. The November 21 confirmation in #542 names an iPhone 17 Pro, iOS 27.3 and a paid developer team.

Can I compile and install Poki directly through Xcode?

Yes: Joseph documents building on a Mac, signing with your Apple Account and installing over USB. Free accounts use PokiExternal, and the instructions call for Xcode 27.1 or later with support for the phone's iOS version. The source-install answer in #542 also covers XcodeGen, Rust, bundle identifiers, App Groups and signing targets. It is a native installation workflow rather than the npm check described above.

Why does Windows installation fail at provisioning profiles after the iPhone has trusted the computer?

The reporter in #421 traced that symptom to stale Windows pairing records. After backing up and clearing the Apple Lockdown folder, reconnecting and accepting Trust, Apple Devices stopped offering its Trust button and installation completed. The confirmed Windows workaround includes closing both Apple Devices and Poki Setup before changing those records and preserving a backup of the folder.

Does Poki work on iOS 18.7.7 after an iPhone restart?

Issue #124 reports a developer-image mount failure after restart on Poki 0.2.9 (672), including unsuccessful USB re-pairing attempts. The runner reached mounting_ddi and reported a device socket I/O failure. The restart report lists repeated connection changes, reinstallations and a Location & Privacy reset among the attempted remedies. This is the documented version for that failure, rather than a reproduction on 0.4.1.

Why do tasks hang after Poki switches to another app?

Maintainers discussed unreliable audio recovery checks in the Poki 0.3.0 background-hang thread, then rolled back the release. Joseph describes 0.3.1 as the 0.2.9 code with the Chinese UI added. Issue #165 also records a completed WeChat task in 89 seconds with 11 tool calls after restarting the session, followed by another occurrence of the hang.

Can I attach workflow JSON, Markdown or SKILL.md files in Poki chat?

Joseph announced JSON, Markdown and plain-text attachments in Poki 0.4.2 Beta with a 100 KB per-file limit. ScottH1122 described collapsed file chips that expand to show contents and text retained through the ordinary message path. The attachment discussion in #369 identifies the supported extensions as .json, .md, .markdown and .txt, with errors for oversized files or invalid UTF-8.

Should a .skill package be read as chat content or imported through a separate skill flow?

Issue #369 leaves skill-package import as a separate design decision. Its shipped attachment change concerns reading JSON, Markdown and plain text in chat, including a document named SKILL.md when handled as Markdown. The package-import discussion separates that question from the attachment implementation. Reading document text and importing a package are different operations in the thread.

What must be verified before the iPhone app reaches version 1.0?

Issue #53 asks for a recorded real-device look, tap and voice acceptance run, first-install qualification and the proposed user-owned Cloudflare Realtime-token endpoint. Joseph lists those items separately from merges already included in Alpha 0.2.9 build 672. The integration checklist also distinguishes phone-side OpenAI token creation from that relay endpoint; it gives acceptance work rather than a release date.

Does the autocomplete query about Poki versus CrazyGames refer to the gaming portal?

The Poki-versus-CrazyGames page discusses browser-game portals. It compares a different entity from JosssphZhou/poki, whose README describes an iPhone agent installed through a Mac or Windows PC. The gaming comparison belongs to that browser-game question. The agent's relevant alternatives here are installation routes and automation or testing tools, as shown in the comparison table above.

What we have not tested yet

Our independent checks cover the Node repository and an aborted deployment attempt, without a completed iPhone installation or task. First-time setup, seven-day signing renewal, cellular cold starts, current Beta background recovery, English app operation and iOS 18 restart recovery remain outside those checks. The deployment attempt has no screenshots or completed known-answer phone tasks. Issue #165 says the exact hang had not been retested on 0.3.1, and #124 has no demonstrated mount repair.

The calculator, WeChat and contributor installation outcomes are attributed reports. Our checks do not include the food orders, rides or flight bookings promoted on X, continuous Screen History, payment and deletion approvals, Wallet protection, attachment privacy in transit or model-provider retention. Windows Apple sign-in lacks the later confirmation available for the Mac and iPhone paths in #542. Its successful November 21 report names one device, iOS version and paid developer team.

The verification guide limits its physical-phone qualification to the Chinese interface in portrait orientation and documents dependence on private Apple interfaces. Its older measurements and protocol checks are not acceptance results for every standalone-app build. The Siri positioning has no independent comparative benchmark here, and the support-sharing guidance does not demonstrate removal of identifiers from archives. The candidate cellular results likewise do not establish that every released build includes those changes.

The audit summary omits the affected package and dependency path, while the test tail omits the failing case needed for a root-cause diagnosis or exhaustive passing-test count. The configured image label does not identify actual Node or operating-system versions; those require raw toolchain logs. Repository activity and issue follow-ups give no adoption count, support guarantee or unattended reliability measurement. A useful next step is the README's disposable Notes task, followed by inspection in Notes.

Alternatives

ProjectWhat it isPick it when
Appium gh↗An open-source WebDriver automation framework with platform drivers and test clients.Choose Appium when you want to write app automation through test clients and platform drivers.
MaestroA UI-testing framework with YAML flows and an MCP interface; its README excludes physical iOS devices.Choose Maestro for supported testing targets, including iOS simulators. Its README excludes physical iOS devices.

What people are saying

  1. [velocity-scout] JosssphZhou/poki

Sources

  1. Poki README
  2. Poki verification and known limits
  3. Poki cellular support discussion
  4. Poki Apple sign-in fix
  5. Poki iOS 18.7.7 restart failure
  6. Appium README
  7. Maestro README

More automation reviews