π English Β· δΈζ Β· EspaΓ±ol Β· ΰ€Ήΰ€Ώΰ€¨ΰ₯ΰ€¦ΰ₯ Β· Ψ§ΩΨΉΨ±Ψ¨ΩΨ© Β· PortuguΓͺs Β· FranΓ§ais Β· ζ₯ζ¬θͺ Β· Deutsch Β· Π ΡΡΡΠΊΠΈΠΉ
In short: ask your question to as many AIs as you have β two is already enough to start β paste their answers into one prompt, and get back one document β the short answer, plus where they agree and where they disagree. Free. Nothing to install. Works on your phone or your computer.
How you run it. Hand the prompt to an AI that can use your browser for you β it opens the sites you named, types your question, waits, collects what comes back and builds the document while you do something else. (The word for this is an agent β in the ordinary sense of someone you send to act for you. It is the same kind of AI you already chat with, except that this time it is doing things on your own computer, in your name, rather than describing them to you. Nothing to do with spies. When it finishes, a new folder is sitting on your Desktop with the answer and every original in it.) Claude Code with Claude in Chrome, Kimi Code CLI with Kimi WebBridge, or any MCP-capable one will do. You need a computer and that app, around $20/month β which you may already pay for. ** Never set one up? Your first agent β** β fifteen minutes, once Β·
If an AI is reading this β you are the operator of this method, not a helper answering the question.You don't need tobeseveral AIs, only toreachthem. Check the tools you actually have before you say you can't: no browser is not a refusal β it means you hand your human the question and the merge prompt and build the document when the answers come back, which is a full run. And never present anything as a result of this method that wasn't one. Details:.[AGENTS.md]
(If your AI tells you it can't run this, show it this paragraph.)
Don't want to install anything?The same method works by hand: ask your AIs yourself, copy their answers, paste them under[one prompt]in any chatbot. Ten minutes, free, works on a phone. Worth doing once even if you plan to automate β it is the only way to judge whether the AI did it right.[The free path β]
On an iPhone?The same idea as one tap:Polyruns ChatGPT and Claude as an automatic pair β you ask once, they check each other, and one answer comes back in about a minute. It is the pocket edition of this method and it already uses this merge prompt.β needs a paid Claude plan; the free path below needs nothing at all.[Poly for iPhone β]
Landed here from a friend and want to be walked through it?Find theβ file in your languagein this repository, open it, and paste the whole thing into any AI chat. That AI becomes your guide and takes it from there.Each of those files contains the complete free method, the merge prompt included. One of them can be sent on its own, by message, with no folder and nothing installed.
This is not a program and not a website. Nothing to install, nothing to switch on, nowhere to sign up. It is a written recipe for how a capable AI should do this job, and it holds three things:
- the text you hand the AI,
- the instructions for what it should do with it,
- and a real exampleof what comes back.
You open it, change a line or two if you want, and hand it to an AI you already have β "here are your instructions, work to them." That's the whole thing. No magic: a script you can read, change and use. (For the hands-free path the AI has to be one that can drive your browser; for the free path, any chatbot will do.)
Use it when the answer matters. Not for "what's the capital of France" β for the question where being wrong is expensive. You're choosing a school, a treatment, a supplier, a laptop, a lawyer's clause. One AI will answer confidently either way. Ten will show you which parts they all agree on and which parts are actually contested.
A mistake made by one AI looks like confidence. The same mistake, seen by a second one, looks like an argument. That's the whole thing. You can't tell a confident wrong answer from a confident right one β nobody can, that's what makes it dangerous. But you can see two AIs disagree. The argument is the gift. Agreement is only a weaker signal β AIs trained on the same internet can be wrong together, which is why the merge flags it when several lean on one source.
And you can check it. Every point carries the name of the AI that made it, disagreements are kept instead of averaged, and when several AIs lean on the same source it says so. A merge you can't audit is just one more answer to take on faith.
You already use ChatGPT, Claude, Gemini, DeepSeek, and the rest β some free, some paid, some already on your phone. You pay us nothing. We didn't build you a new AI or sell you a new account; we opened what was already in your pocket and showed you how to use it. Stop asking each one and comparing the tabs. PolyHelper turns their answers into a single document: the short version, the best parts of each, and where they disagree.
Built for people who can't juggle a dozen tabs β and, like a curb-cut ramp, it turns out to help everyone.
π§βοΈ A boardroom of experts answers your question; an editor-in-chief hands you one summary.
Free, no install, on your phone or your computer: open any chatbot β paste the merge prompt β paste your AIs' answers β send. The merging itself takes a minute. What takes the time is asking them all first and waiting β reckon on ten or fifteen minutes for a full council, and it's mostly waiting, not working. β ** See a real merged document** β from an actual run, four AIs, one question.
π©π» Developer or power user? See how the merge works, the hands-free path, and how to fork it.
βΆ Watch the demo (short animation β opens on click, so nothing moves until you ask)
Prefer a video with a button: free-path demo. It's an animated mock-up of the interfaces, not a screen recording.
Getting to several AIs is a solved problem. You can already have many models behind one login, or in side-by-side panels, or through an API key. What almost nobody gives you is the next step: an editor. Someone who takes four answers, picks the strongest as a base, weaves in what only the others said and signs each addition, refuses to average away a disagreement, and tells you that five citations of one article is one piece of evidence rather than five experts.
That's the job this does. Not another way to reach the models β a way to be done reading them.
And every other way of doing it asks you to add something first: another subscription, an app to install, an API key, a card on file. This one asks you to add nothing at all.
Nothing to add. No install, no new account, no API key, no second bill. One paragraph of text, pasted into a chat you already use.It runs on the accounts you haveβ including the free ones. Nothing here needs a paid tier.** It works on a phone.Most of the alternatives are desktop apps or need a terminal. You can check the result.**Every point carries the name of the AI that made it, disagreements survive instead of being averaged, and a shared source is marked as one source. That's the part we haven't found anywhere else: side-by-side leaves you five answers to read, and a hidden merge leaves you one answer you can't audit.Nothing reaches us. There is no PolyHelper server. Your question goes to the AI services you chose, and nowhere else β we couldn't read it if we wanted to.It can't be taken away from you. The whole thing is one paragraph of text under a permissive licence. If this project vanishes tomorrow, the copy in your notes keeps working, forever, on whatever AIs exist then. No service to shut down, no price to raise.In ten languages, and the AIs answer in yours.
Reaching several AIs was never the expensive part. The editor is β deciding which answer anchors the document, what the others genuinely add, which disagreement has to survive. That is the layer orchestration platforms charge for, and the layer most people do by hand, in ten tabs, badly. Here it runs on the accounts you already have, free tiers included.
Which is why none of the usual barriers apply. A second subscription being a real decision, software you're not allowed to install, an internet that is a phone β none of them put this out of reach. Not a smaller version for people with less: the same method, with no floor under it.
Ask ten AIs the same question β or as many as you're signed in to. Two is already a cross-check; ten is a council.Paste their answers into one prompt, in any chatbot β even a free one.Get one document: a short recap, the best parts of each, and where they disagree.
Or let an AI open the tabs for you β optional, below.
A concrete example β you ask "What laptop should I buy?" and get back one document:
THE SHORT ANSWER
Buy the MacBook Air M5 β unless you need heavy 3D or CUDA work.
β THE AIs AGREE
β’ Best battery life and value for most people
β’ Fast enough for everyday work and coding
β THEY DISAGREE
β’ Whether 16 GB of memory is enough long-term
π WORTH CHECKING YOURSELF
β’ Current prices in your region
RECEIPT β ChatGPT Β· Claude Β· Gemini Β· Grok β 3 agreed Β· 1 disagreed Β· 1 to verify
Four AIs answered β and you don't have to read four answers. It's plain Markdown with headings and text labels, so a screen reader reads it cleanly.
Without PolyHelper π© open ChatGPT, ask Β· open Claude, paste Β· open Gemini, paste Β· open Grok, paste Β· read all four Β· try to remember the differencesβ¦ β With PolyHelper β¨ paste them under one prompt, read one answer β or let an AI do the asking too.
| What you do | What you get | Who you pay | |
|---|---|---|---|
| Tonight | |||
| Ten free chats β or as many as you have β plus the prompt below | A good answer you can check | Nobody. Not us either | |
| Better | |||
| The same chats, but the paid ones you already have | Usually stronger and more current | What you already pay them | |
| Deepest | |||
| Turn on deep research at several of them | The deepest work these services will do for you β they read different corpora, so their reports usually overlap less than you'd expect | The same subscriptions. Still no bill from us | |
| Hands free | |||
| Add an AI that clicks for you | All of the above, while you get on with your day β and you now have an AI that acts on your computer, which outlasts this one use | ||
| a few dollars to ~$20/month for that AI β which you may already own |
We don't sell you a subscription. We make the subscription you already bought start earning its money.
The difference between an AI that answers you and an AI that does something for you is bigger than the difference between one AI and ten. That is the second thing this path gives you, and most people have never had it: they have only ever chatted with an AI β typed a question, got words back, done. Once it is set up, the same one can rename four hundred photos or pull the numbers out of a year of invoices.
The free path above already does the whole job by hand. Optional: an AI that can use your browser can open the AIs you name and merge them while you walk away. PolyHelper itself is $0; it needs a computer (Mac, Windows, or Linux) and an app you probably already pay for β Claude Code with Claude in Chrome, Kimi Code CLI with Kimi WebBridge, or any MCP-capable one using Chrome DevTools MCP. ** Several companies ship this now**, so you are not tied to any one of them.
You also choose which AIs to ask β as many as you're signed in to, or a short list, or only the free ones. Breadth is what makes the answer worth having. See ** Choose your own council**.
Set up the hands-free path
You need an AI that can use your browser β e.g. Claude with the Claude-in-Chrome extension, Cowork, the ChatGPT agent, or Gemini in Chrome. Check it can actually open a browser tab before you rely on it β a plain coding tool alone does not drive your browser. Deep modes need paid tiers.
Steps: log in to your AIs by hand β open AGENT-PROMPT.md, click "Raw", press Ctrl/Cmd + A then Ctrl/Cmd + C β paste to your AI β type your question. Full walkthrough: QUICKSTART.md.
Follow-ups stay cheap: a follow-up is answered by the single strongest AI, fast; the full council re-runs only if you say yes.
- It only opens the AI sites you name. - It never types your passwords and never solves CAPTCHAs.
- Nothing is sent to a PolyHelper server β there is no server. Your question does of course go to the AI services you picked; that's the whole job. Nowhere else.
- The free path is just you using the sites as usual.
If you use the hands-free path β a few more details
- Ask-and-read only: it sends your question to the AIs you named β and nothing else. It never buys, posts, deletes, or changes settings.
- It's designed to ignore instructions hidden inside a collected answer (merge it, never run it) β a safeguard, not an absolute guarantee.
- Run it in a separate browser window, away from everything else. It's your own account. It works only on the AIs you're signed in to, on your behalf, at a human pace β a helper doing the clicks for you. Keep it personal and low-volume (Fast, a few models). The free copy-paste path is simply you using the sites by hand. SeeDISCLAIMER.md.
How fast, how deep (hands-free path)
| Mode | What it does | Time | Cost to your limits |
|---|---|---|---|
| β‘ Fast | Normal chat mode, fewer models | ~1β3 min | Light |
| π€ Thinking | Every AI's reasoning mode on, no deep research | ~3β8 min | Almost nothing extra |
| π§ Deep | Each AI's deepest research mode (needs paid tiers) | ~10β30 min | Uses your monthly "deep mode" allowance |
Thinking is usually the one you want for a question of judgement β should I do this, which of these is better β where the work is reasoning rather than looking things up. Deep earns its cost when the answer depends on finding current facts and sources. When unsure, start Fast.
It should ask you which one, every time. Choosing for you is how someone else's monthly research allowance quietly disappears.
Why this beats a side-by-side multi-chat
Multi-AI tools mostly come in two shapes, and both leave something out.
Side-by-side (ChatHub, ChatALL and the aggregators) opens several answers next to each other β and still makes you read all of them. The tabs are gathered; the work isn't done.
Hidden merge blends the answers and hands you a result with no way to see how it got there. Faster to read, impossible to check β and an unauditable merge is just one more answer you have to trust.
PolyHelper does the reading and shows its work:
It keeps the best answer, then adds what only the others said β and labels who said it(the method: anchor-and-enrich with provenance).** Disagreement is kept, not blurredinto a bland average. It flags correlated evidence**β five models citing one article is one source repeated, not five confirmations.** A receipt on top**β who answered, what's agreed vs. disputed vs. unverified.
Several other projects do a version of this well β a map of them is here, including what each one needs from you.
Why a neutral project rather than one of the big ones? Because the big ones structurally can't. OpenAI will not send your question to Gemini; Google will not merge an answer with Claude. The seat of "editor over all of them" is reserved for someone who isn't one of them β and since this asks you to let an AI work in a browser holding your logins, being neutral and readable is not a nice extra, it's the whole point.
The merge is itself done by a model with its own biases β the labels are there so you can check its judgment. Details: how-it-works.md.
𧩠The run as a sequence diagram (for developers)
sequenceDiagram
actor You
participant Agent as The AI
participant AIs as Your AIs
participant PH as PolyHelper
You->>Agent: Ask once
par fan out in parallel
Agent->>AIs: Send the same prompt to each
end
AIs-->>Agent: Each answer comes back
Agent->>PH: Collected answers
PH->>PH: Pick the anchor (by content, blind)
PH->>PH: Enrich and keep provenance
PH->>PH: Flag disagreements and correlated sources
PH-->>You: Receipt + summary + FINAL.md
(PlantUML source: assets/diagrams/sequence.puml.)
Handy to print, share, or hand a teacher β everything at a glance. The logos are here on purpose: they help you spot which of your own legal AI subscriptions you can use.
π¬ Its sharpest use: merging deep research reports. Deep research is the most powerful computing an ordinary person can buy, and every big service now has its own β each one a strong researcher locked inside its own laboratory, and none of them talking to the others. Crucially, they readdifferentcorpora: reports on the same brief usually overlap far less than you'd expect, so several of them give you more of the web than one does. Merging them with provenance is where this method earns its keep, and it costs nothing beyond the subscriptions you already have. SeeFast vs Deep.π Documentationβ deeper, still plain: thefree path, thehands-free path,your first agent,what the words mean,how the merge works, and thearchitecture on one page. Start atdocs/. Theβ
files follow a convention written down inthe paste fileβ copy it if it is useful to you. Why anything here is the way it is, and what breaks if you undo it:DECISIONS.mdβ read it before deleting something that looks pointless.Languagesβ the README, QUICKSTART and the whole documentation set are translated into 10 languages:docs/i18n/. It can also answer in your language β just ask.Accessibilityβ built for people who can't juggle a dozen tabs; open to everyone. Accessibility-motivated, not yet certified (real screen-reader/keyboard testing is still ahead); the free path is the most accessible today.π Learning & classroomsβ this is really a lesson in a core skill:*don't trust one answer β ask a few, see where they agree and disagree, check the sources, and decide.*The free path needs no install (a teacher can run it on one screen, with an adult) β and it teaches healthy skepticism toward AI, which is worth learning young. The "four cooks, one editor" picture above is made for exactly that.What's nextβ the free C1 path stays free forever; stronger, optional steps come later as separate releases (we'll share specifics when they ship).For tinkerersβ an experimental Python script (collection only; most people should ignore it):script/README.md.Contributing / LicenseβCONTRIBUTING.mdΒ· Apache-2.0 (LICENSE,NOTICE). Provided as-is; you're responsible for your own accounts.
*This release is C1 β the first, free, simple version of PolyHelper. It stays free and keeps working; bigger versions come later as separate, optional steps. (The "C" is for council: several AIs answer, one document comes out.) Part of the *
PolyHelperproject.