Two AIs work on your question as a pair and check each other, on your iPhone. It runs on the ChatGPT and Claude subscriptions you already pay for — no API keys, no tokens to buy, no server, no account to create.
Orchestrating several models usually means paying twice: once for the subscriptions you use every day, and again for API calls so those models can work together. Poly does it on what you already have.
🌍 English · العربية · Deutsch · Español · Français · हिन्दी · 日本語 · Português · Русский · 中文
Part of the PolyHelper family — the method, in your pocket.
C1M is the project; Poly is the shortcut you install on the phone.
Tap Poly → type your question → pick Critique → ~1 minute → the answer is on screen and in your clipboard
- Download
Poly.shortcutfrom the latest release. - Tap the file on your iPhone → Add Shortcut.
- Run it once and tap Always Allow when it asks for ChatGPT, Claude and clipboard access.
Requirements: iPhone on iOS 18 or later, the ChatGPT and Claude apps installed and signed in, and a paid Claude plan. Full walkthrough: docs/en/install.md.
One model answering alone is confident about everything, including the parts it gets wrong. A second model reading the first one's answer catches factual slips, missing angles and overconfident claims — and it costs one extra message.
Poly runs on the Shortcuts actions the ChatGPT and Claude apps publish themselves, so a run spends messages from your own plan, exactly like asking the question by hand.
Use it when the answer matters. Not for "what's the capital of France" — for the question where being wrong is expensive: a contract clause, a diagnosis you're about to act on, a supplier, a treatment, a piece of code going to production, a letter you can't unsend. People already do this by hand — asking one AI, then another, then squinting at two tabs. Poly is that habit without the tabs.
And it keeps the argument. When the two models disagree, Poly says so instead of averaging them into one smooth paragraph: what each one claimed, which part is checkable, what to verify. Two AIs agreeing is a weak signal — they read the same internet — so agreement is never presented as proof.
your question
│
├──→ ChatGPT drafts the answer
│
└──→ Claude verifies it and fixes what's wrong
│
└──→ one final answer → screen · clipboard · journal
That's Critique, the default mode and the only one you need on the first day. Ten other modes put the two models against each other in different ways — a debate, a side-by-side, a map of where they disagree — and they stay out of your way until you want them.
All eleven modes
| Mode | What happens | Cost |
|---|---|---|
| ⚖️ Critique | ChatGPT answers, Claude verifies and delivers the improved final. Your default. | 2 messages |
| 🩺 Advisor | Claude reviews your own text without rewriting it: verdict, strongest objection first. | 1 message |
| 👀 Side by side | Both answer independently, answers shown next to each other, nothing blended. | 2 messages |
| 🔀 Synthesis | Both answer blind, then one answer anchors the assembly and the other adds only what's missing. | 3 messages |
| 🧭 Auto | Not sure which mode fits? One classification call recommends one. | 1 + the mode |
| ⚔️ Decision | Fast view vs cautious view, then a referee: first steps, risks, confidence. | 3 messages |
| 🗺 Dispute map | Where the two agree, where they disagree, what is checkable — no verdict imposed on you. | 3 messages |
| 🥊 Debate | Draft → opponent attacks it → revision → judge. For the hard ones. | 4 messages |
| ❓ Clarify | The model asks you what's missing first, then answers. | 2 messages |
| 🎨 Image | Both draw the same brief blind as vector art; you get a gallery and pick. | 2–3 messages |
| ➕ Delta | Claude anchors, ChatGPT returns only what it would add. A cheap stand-in for Synthesis. | 2 messages |
| ℹ️ What is Poly | Built-in help, right inside the run. | free |
Five everyday modes sit in the main menu; the rest live one tap away under 📂 More… The ✉ badge on each entry shows the price before you choose. Which mode for which task: docs/en/recipes.md.
Small single-purpose shortcuts that share the same engine:
- Poly Voice — tap, dictate, hear the final answer read back.
- Poly Photo — share a photo or PDF; the text is recognized on device, then answered.
- Poly Quiet — the Critique pipeline with no screens at all: clipboard and journal only.
- Poly Compress — shrink a long text with Apple's on-device model before the pair sees it. This one needs Apple Intelligence hardware (A17 Pro / M-series or newer); nothing else does.
- Poly Multi — the pair plus any other AIs you have, semi-manually: Poly writes the prompt, walks you through the apps and merges everything into one document; you paste, send and copy. Roughly half the work of doing it by hand. Ships alongside the core.
Want the full council rather than a pair? Poly is the pocket edition of a larger method: PolyHelper asks up to ten AIs and merges them into one document with the disagreements kept — free, nothing to install, on any device. Poly Multi already merges by its canon. → PolyHelper/C1
Poly proper is an iPhone shortcut. But the same idea runs on Android through Tasker and the clipboard, because neither Claude nor ChatGPT publishes an action there that another app can call. Tasker prepares the prompt, opens each app in turn and picks the answer up the moment you copy it; you do the tapping, it does the choreography and the merge.
It is built and structurally verified, and nobody has run it on a real phone yet. That is the honest status, and it is exactly where you can help: import the clipboard relay, ask it one question, and tell us what happened — android/ has the files and the guide.
English is the base language. Everything a user sees — menus, notifications, prompts, journal
entries, docs — comes from locales/<lang>/, so a new language is a folder, never a fork of the
code.
Prompts are deliberately shared across all languages. The models read them, not you, and English
works best there — every locale uses locales/en/prompts/ by default, and any language may
override a single prompt by dropping its own file into locales/<lang>/prompts/ if it ever needs to.
All ten are built and verified against their sources — python3 tools/verify.py checks that every
shipped shortcut is what the locale files produce, and that every kit matches.
| Language | Locale | Language | Locale | |
|---|---|---|---|---|
| English (base) | en |
日本語 | ja |
|
| العربية | ar |
Português | pt |
|
| Deutsch | de |
Русский | ru |
|
| Español | es |
中文 | zh |
|
| Français | fr |
हिन्दी | hi |
Verified against the source files is not the same as run on a phone. Two of the ten have been through a live iPhone with paid accounts, end to end, every mode; the other eight are built from the same sources and machine-checked against them, but nobody has used them on a phone yet — if you install one, telling us what broke is the most useful thing you can do.
Adding a language, or fixing one: CONTRIBUTING.md.
The answer always comes back in the language of your question, whichever locale you install — the prompts instruct both models to follow the user's language.
Your question goes to the two apps you already use, and the result stays on your phone. There is
no C1M server for it to reach — nothing is collected, and there is nothing to sign up for. The
journal (Poly-journal.md) is a plain text file in your own iCloud Drive; delete the journal
action or the file if you don't want a history. A paid iCloud plan is not required — without
iCloud the answer still reaches your screen and clipboard, only the journal is skipped.
Message consumption comes out of your ChatGPT and Claude subscriptions, 1 to 4 per run, from the same allowance as your normal chats — and the ✉ badge in the menu tells you the price before you pick a mode. Poly is open source under Apache 2.0 and free to use. Some future companions or hosted conveniences may be offered separately.
Worth knowing before you install:
- iPhone only, iOS 18 or later. iPad is expected to work on iPadOS 18+, but it is untested.
- A paid Claude plan is required. On a free account the Claude action can come back with "model isn't available."
- A run takes about a minute, closer to two in the multi-step modes.
- It spends messages from your subscriptions, not from an API budget.
- Touching the screen cancels the run. That's a limit of Apple's platform, not a choice we made. The one exception is Clarify, which opens its own window and asks you to answer.
- You can't pick the model from inside Poly. Each app uses whatever is set as its default, so check the model in the Claude app before a run that matters.
- Very long questions are risky. The Claude action can time out and hand back an empty result while Claude keeps writing inside its own app; shorter questions return reliably.
The demo above is a rendering of the interface — the labels come from the English locale and the answer in the last frame is the real output of the run shown, copied from the phone. All eleven modes were run end to end on a live iPhone with paid ChatGPT and Claude accounts — both menu levels, the finale menu, the result screen, clipboard delivery, the journal and the permission flow. That was one phone in two of the ten languages; the other eight are built from the same sources and machine-checked against them, and the Android path has never been run at all. Everything known to be unverified is in Limitations above and in the FAQ.
./tools/build.sh # every locale: build, sign, package, verify
./tools/build.sh en # just one
python3 tools/verify.py # checks onlyRequires macOS (the shortcuts CLI signs the files) and Python 3.9+. The builders read
locales/<lang>/ and write dist/<lang>/; tools/verify.py unpacks the signed result and
checks that it matches what the sources produce.
Install, step by step · What Poly is, in plain language · Recipes: which mode for which task · FAQ · Roadmap · Changelog · Security policy · Contributing
The most useful thing you can send is not a star. Read this page for thirty seconds, then tell us whether it was obvious what Poly does, who it's for and how to run it. If it wasn't, open a Discussion and say where you got lost — that's the feedback we can't get any other way.
- Discussions — ideas, questions, proposals for new modes.
- Issues — bugs and concrete tasks, using the issue template.
- Security — please don't file a vulnerability as an issue; follow SECURITY.md.
Poly is an independent open-source project. It is not affiliated with, endorsed by, or sponsored by OpenAI or Anthropic. "ChatGPT" and "Claude" are trademarks of their respective owners, used here only to identify the apps Poly talks to. See TRADEMARKS.md.
Poly is a personal tool — not professional advice and not a fact-checker. Please read the disclaimer before relying on it for anything that matters.
Licensed under the Apache License 2.0.