Reply Wizard
The long answer
Reply Wizard, explained
Everything the short pages skip: how the extension actually reads a chat, every permission it asks for and why, the three ways to run the AI, and a plain list of the things it cannot do. Written to be read by a person and quoted by a machine.
What Reply Wizard is
Reply Wizard is a Google Chrome extension (Manifest V3) that adds AI features to WhatsApp Web at web.whatsapp.com. It transcribes incoming voice messages, suggests replies drawn from the recent conversation, summarises long chats and long voice notes, types by microphone dictation and rewrites a draft you have already written. It never sends a message on your behalf: every result is placed in the WhatsApp message box for you to read, edit and send yourself.
It is made by Palworks FZE LLC, a software company in Ajman, United Arab Emirates. It is an independent product and is not affiliated with, endorsed by or sponsored by WhatsApp LLC or Meta Platforms, Inc.
Reply Wizard is free with no account and no sign-up when you supply your own AI provider API key or run a model on your own machine. Paid plans, at $1.99 or $4.99 per month, exist only so that you do not have to hold a key: they replace the key, not the features.
How an extension reads a WhatsApp chat
Reply Wizard reads a WhatsApp chat the same way your eyes do: from the page. It runs a content script inside the web.whatsapp.com tab and reads the message text that WhatsApp has already drawn on screen. There is no WhatsApp API involved, no login is shared with anyone, and no unofficial client connects to WhatsApp's servers on your behalf.
A content script is a standard browser-extension mechanism: a small file the browser injects into a page you are already viewing, which can then read and change that page. Chrome documents it as files that run in the context of web pages. It is the same mechanism a password manager uses to fill a login form, or an ad blocker uses to hide an advert.
Because Reply Wizard works from the rendered page, three things follow, and they are the whole shape of the product:
- It only knows what is on screen. A chat you have not opened is a chat it has not read. WhatsApp does not hand it your archive, because WhatsApp is not handing it anything.
- It cannot log in as you. Your WhatsApp session belongs to WhatsApp Web. Reply Wizard has no credentials, no session token and no way to reach WhatsApp's servers.
- It cannot press send. Text goes into the message box. Sending it is a thing your hand does.
Source: Chrome for Developers, Content scripts.
What it can see, and what it actually looks at
Any Chrome extension granted access to web.whatsapp.com is technically capable of reading everything that page displays. That is true of Reply Wizard and of every other extension of this kind, and an honest answer has to start there. What separates them is what the extension does with that capability. Reply Wizard reads the recent messages of the one chat you have open, at the moment you press a button, and nothing else.
Specifically:
- Opening a chat sends nothing anywhere. Scrolling, reading and typing are invisible to us. The extension is idle until you ask it for something.
- When you ask, it sends only that request. A reply suggestion sends the recent messages of the open chat. A transcription sends that one audio clip. A summary sends that one thread. Nothing else travels with it.
- It goes to the AI provider you chose, not to us. If you brought your own API key, the request goes straight from your browser to that provider under that provider's terms, and never passes through any server of ours. If you are on a paid plan it passes through our gateway so that usage can be metered, and the message content is not retained after the request completes.
- No chat history is stored on our servers. There is no copy of your conversations anywhere in our infrastructure, because none is ever uploaded.
- The on-device options send nothing at all. Whisper transcription and Chrome's built-in Gemini Nano run inside your browser. When you use them, no message and no audio leaves your computer.
About your API key. It is encrypted with AES-GCM and kept in the extension's own storage on your device, and it is never transmitted to us. Be precise about what that protects against: the encryption key sits on the same device, so this stops casual inspection rather than someone who already has access to your computer. The real boundary is that the key never leaves the machine you typed it on.
Every permission, and why
Reply Wizard requests six Chrome permissions and access to exactly one website. It asks for no access to your browsing history, your tabs, your bookmarks, your downloads, your clipboard or any site other than WhatsApp Web. The full list is below, taken from the extension's own manifest.
| Permission | What it is for |
|---|---|
storage | Keeps your settings, your saved templates and your encrypted API key inside your own browser. |
scripting | Puts the Reply Wizard bar into a WhatsApp Web tab, including after the extension updates. |
offscreen | Runs on-device Whisper transcription in a hidden background document, because speech recognition needs a page context that a service worker does not have. |
unlimitedStorage | Holds an on-device speech model after you choose to download one. Without it, Chrome's default quota is too small for the model file. |
alarms | Fires a reminder at the time you set for it. |
notifications | Shows that reminder when it fires. |
https://web.whatsapp.com/ | The only website Reply Wizard is allowed to touch. Every AI provider host is an optional permission, requested one at a time when you save a key for that provider, and it can be revoked from Chrome's own extension settings. |
Each name above is a standard Chrome permission; Chrome publishes what each one grants in its permissions reference.
One consequence of Manifest V3 is worth knowing, because it is a security property rather than a promise we are making. A Manifest V3 extension may not load and run code from a remote server; Google both requires this technically and enforces it as Chrome Web Store policy. Everything Reply Wizard executes was reviewed by Google before it reached you, and it cannot be quietly swapped for something else between reviews.
Whether it puts your WhatsApp account at risk
Only WhatsApp can say what its terms allow, and we will not pretend otherwise. What we can state precisely is what Reply Wizard does: it does not automate sending, it does not send at all, it does not bulk-message, it does not schedule, it does not connect to any unofficial WhatsApp API, and it does not use a modified WhatsApp client. It reads the page you already have open, in the browser you already logged into, and writes text into the box you were going to type in.
The behaviours messaging platforms usually act against are automated or bulk sending and unofficial clients. Reply Wizard is architecturally incapable of the first: nothing in it can call WhatsApp's send. And it is not a client at all; it is an overlay on the official web app, which continues to do all the messaging itself.
If this matters to you, read WhatsApp's Terms of Service and decide for yourself. We would rather point you at the primary source than summarise it in our own favour.
Three ways to run the AI, compared
Reply Wizard can get its intelligence from three different places, and the choice is yours at any time. Bring your own API key and Reply Wizard is free forever. Run a model on your own computer and nothing leaves the machine. Or subscribe, from $1.99 a month, and we supply the key. The features are identical in all three; only the cost and the path your text takes are different.
| Your own API key | On your own device | A Reply Wizard plan | |
|---|---|---|---|
| What you pay us | Nothing, ever | Nothing, ever | $1.99 or $4.99 a month |
| What you pay elsewhere | Your provider's usage rates | Nothing | Nothing |
| Where your text goes | Straight to the provider you chose, never through us | Nowhere. It stays in your browser | Through our gateway to the model, not retained afterwards |
| Setup | Create a key at the provider, paste it once | Download a model once, from within the extension | Subscribe. There is nothing to configure |
| Choice of model | 13 providers, and any local OpenAI-compatible endpoint | Whisper for speech, Chrome's Gemini Nano for text | Curated. Power adds the larger models |
| Works offline | No | Yes, once the model is downloaded | No |
| Best for | Anyone who already holds an AI key | Anyone who wants nothing to leave the machine | Anyone who does not want to think about keys |
The thirteen providers you can bring a key from are OpenAI, Anthropic, Google Gemini, Mistral, OpenRouter, DeepSeek, xAI, Zhipu, Moonshot, Qwen, Groq, Together AI and Cerebras. A local endpoint, such as LM Studio, is also supported, and for security it must be on localhost: the extension's content security policy fixes the set of hosts it can reach at build time, so a typed remote URL is refused rather than trusted.
How to transcribe a WhatsApp voice message
To read a WhatsApp voice message as text instead of listening to it: install Reply Wizard in Chrome, open the chat at web.whatsapp.com, and press the transcribe control on the voice note. The text appears under the message in a few seconds. You can transcribe in the cloud through a provider key, or entirely on your own device with Whisper, in which case the audio never leaves your computer.
-
Install the extension
Add Reply Wizard from the Chrome Web Store and open web.whatsapp.com. No account or card is needed.
-
Pick a speech engine
In the extension's settings, choose a cloud engine (Groq, OpenAI or Google Gemini, using your own key) or choose on-device Whisper and download a model once. Whisper is offered in several sizes, from
tinyfor a modest machine up tolarge-v3-turbo, and the extension benchmarks your GPU to recommend one. -
Transcribe the voice note
Open the chat and press the transcribe control on the voice message. The transcript appears beneath it, with buttons to copy it, drop it into your draft, redo it, or change the settings for that one message.
Whisper is the open speech-recognition model published by OpenAI; Reply Wizard runs it in your browser through ONNX Runtime and WebGPU. Source: openai/whisper. Chrome's on-device text model, Gemini Nano, is documented as Chrome built-in AI.
How to summarise a long group chat
To catch up on a WhatsApp group without reading every message: open the group at web.whatsapp.com with Reply Wizard installed and press Summarise. The extension reads the recent messages that are on screen and returns a short account of what was decided, what was asked and what is still open, in a panel over the chat. A long voice note can be summarised the same way, from its transcript.
-
Open the group
Scroll back as far as you want the summary to cover. Reply Wizard summarises what the page holds, so how far you scroll is how much it reads.
-
Press Summarise
The control sits in the Reply Wizard bar under the message box. How many messages are included is a setting you control.
-
Read, then act
The summary opens in a panel. Nothing is posted to the group: a summary is for you, and the group never learns that you asked for one.
What Reply Wizard cannot do
Reply Wizard cannot send a message, cannot auto-reply, cannot bulk-message, cannot schedule a message, cannot run on the WhatsApp phone or desktop apps, cannot run on Telegram, Slack, Messenger, Instagram or Discord, cannot read a chat you have not opened, and cannot work without WhatsApp Web open in front of it. Several of these are the point of it rather than gaps in it.
- It cannot send. There is no send path in the extension at all. Reminders open a chat and stop; templates fill the box and stop.
- It cannot auto-reply. Nothing happens unless you press something. There is no bot, no trigger and no rule engine.
- It cannot bulk-message or schedule. Neither exists, and neither is planned.
- It does not run on mobile. Chrome extensions do not run on the WhatsApp iPhone or Android apps, and Reply Wizard is a Chrome extension.
- It does not run on other messengers. WhatsApp Web only. If that changes we will say so here, and not before.
- It cannot read a chat you have not opened. No archive access, no background scanning, no crawling your chat list.
- It does not keep your conversations. No chat history is stored on our servers, because none is uploaded to them.
- It cannot reach a remote custom endpoint. A self-hosted model has to be on
localhost. This is a deliberate restriction, not a missing feature.
The words used on this site
Five terms appear throughout Reply Wizard's documentation and are worth defining once: BYOK, token, on-device, Manifest V3 and MCP.
- BYOK, or bring your own key
- Using Reply Wizard with an API key you obtained yourself from an AI provider. The provider bills you directly at their rates, Reply Wizard charges nothing, and your requests never pass through our servers.
- Token
- The unit AI providers bill in, roughly three quarters of an English word. Reply Wizard's Pro plan includes 2.5 million tokens a month, which is about 2,500 reply suggestions; the Power plan includes 10 million.
- On-device
- AI that runs inside your own browser rather than on someone's server. In Reply Wizard this means Whisper for speech-to-text and Chrome's built-in Gemini Nano for text. Nothing is uploaded, and it keeps working with no internet connection once the model has been downloaded.
- Manifest V3
- The current Chrome extension platform. Its most relevant rule here is that an extension may not download and run code from a remote server, so what Google reviewed is what runs on your machine.
- MCP, or Model Context Protocol
- An open standard that lets an AI assistant read from a tool. Reply Wizard can expose your WhatsApp Web to an assistant on your own computer through it. The connection is local-only and read-only, it is off until you turn it on, and your API keys never travel across it. See modelcontextprotocol.io.
Questions
More questions
The eight most common ones are answered on the home page. These are the rest.
Does Reply Wizard work on the WhatsApp mobile app?
No. Reply Wizard is a Chrome extension, and Chrome extensions do not run inside the WhatsApp app on iPhone or Android, nor inside the WhatsApp desktop app. It works at web.whatsapp.com in a Chromium desktop browser, and nowhere else.
Can I use it without an internet connection?
Partly, and this is the one case where the answer is genuinely yes. If you have downloaded an on-device Whisper model, transcription runs with no network call at all. Chrome's built-in Gemini Nano is on-device too. Anything using a cloud provider needs a connection, and WhatsApp Web itself needs one to show you the chat.
Which languages does it work in?
Microphone dictation offers Chrome's own speech-recognition set, 101 language tags covering the major world languages and their regional variants, and you can shortlist up to five and switch between them from the reply bar. Replies, summaries and rewrites work in whatever languages the AI model you selected supports, which for the major providers is most of them.
Does Reply Wizard slow down WhatsApp Web?
It should not. The extension is idle until you press something: it adds a bar under the message box and then waits. The one thing with a real cost is on-device transcription, which uses your GPU while it runs, and it only runs when you ask for a transcript.
Can WhatsApp, or my employer, see that I am using it?
Reply Wizard sends nothing to WhatsApp and announces itself to nobody. It does not post, mark or watermark anything, and a summary or a transcript is never visible to the other person in the chat. We cannot make claims about what a managed device or a corporate browser policy can observe about installed extensions, because that is a property of your machine and not of this software.
What happens to a transcript after it is made?
It stays in your browser so that you can reopen the same voice note without paying to transcribe it twice. Audio captured for a transcription is held only for as long as that transcription takes and is not retained by us afterwards. Nothing is stored on our servers at any point.
Can I point it at my company's own model endpoint?
Only if that endpoint is on your own machine. Reply Wizard supports any OpenAI-compatible server running on localhost, which covers LM Studio, Ollama behind a compatible shim, and similar. A remote URL is deliberately refused: the set of hosts the extension may contact is fixed when it is built, so a typed address cannot become an exfiltration route.
Does it work in group chats?
Yes. Summaries are the feature groups benefit from most, and reply suggestions read the recent group messages the same way they read a one-to-one thread. Reminders behave differently in a group: rather than opening the chat with text prefilled, they show you a card, because prefilling a message into a group is a mistake waiting to happen.
Do you train AI models on my messages?
No. Palworks trains nothing and stores no conversations. When you use your own API key, what that provider does with the request is governed by that provider's terms, which you should read, and several of them let you opt out of training. When you are on a Reply Wizard plan, the request passes through our gateway to a model and the content is not retained after it completes.
What happens when my subscription ends?
The extension keeps working. A lapsed or cancelled plan falls back to bring-your-own-key and to the on-device options, both of which are free and neither of which is degraded. You lose the managed key, not the software.
Is Reply Wizard open source?
No, the source is not public. What is verifiable is that every Manifest V3 extension ships its code to your machine and cannot fetch more at runtime, so what Google reviewed is what runs. If you want to read what a version does, the code is on your own disk after you install it.
How do I uninstall it, and what is left behind?
Remove it from Chrome's extensions page like any other extension. Chrome deletes its storage with it, which includes your settings, your saved templates, your API key and any on-device model you downloaded. There is no account of ours to close, because you never made one.
Read all of that and still curious?
The free version needs no account, no email address and no card, so trying it is cheaper than deciding.