Privacy Policy

Last updated: October 18, 2026

The short version

This policy has two main parts: The website and The app. They work very differently, so we describe them separately.

Who we are

talkflow is run by Samir Kharel, an individual developer ("we", "us", "our"). You can contact us at samir@talkflow.live.

For the website information described in Part 1, and for any emails you send us, we decide how that information is used. Under data protection laws such as the GDPR and UK GDPR, that makes us the "controller" of that information.

What this policy covers, and what it doesn't

This policy covers:

It doesn't cover:


Part 1: The website

What information is collected

When you visit the site, your browser sends our hosting provider the technical information that every website needs in order to respond. This typically includes:

Our hosting provider is Vercel Inc.

What the site doesn't do

The GitHub star count

The page shows how many people have starred the project on GitHub. Our server fetches this public number from GitHub's API, at most once an hour. Your browser does not contact GitHub for this, and GitHub does not receive your IP address from it.

Why we use it, and our legal basis

This information is used only to deliver the site to you, keep it running, and protect it from abuse such as attacks and excessive traffic.

Under the GDPR and UK GDPR, our legal basis is our legitimate interests in running a working and secure website (Article 6(1)(f)). You have the right to object to this. See "Your rights" below.

Who handles it for us

Our hosting provider (Vercel Inc.) handles this information in order to serve the site. See Vercel's Privacy Notice.

The app's one-time install request (see What we receive from the app) reaches us the same way, through Vercel. The total it adds to is kept in a database hosted by Neon, which receives only that number.

We don't share website information with anyone else, except where the law requires it.

How long it's kept

Vercel keeps request logs for a limited time under its own policies. We don't copy or export them.

Downloads

If you download or view talkflow on GitHub, GitHub receives your request, including your IP address, and handles it under the GitHub Privacy Statement.

"Do Not Track" and Global Privacy Control

We don't track visitors across other websites, and we don't allow other companies to do so through our site. Our aggregate analytics don't use cookies or build a profile of you. So there's nothing for a "Do Not Track" or Global Privacy Control signal to switch off. The site works the same way whether or not your browser sends one.


Part 2: The app

talkflow has a Mac app and a Windows app. They handle your information the same way; where they differ, this part says so.

How dictation works

When you hold the talkflow shortcut (fn on a Mac and Ctrl + Win on Windows, unless you change it), the app:

  1. records sound from your microphone and keeps it in your computer's memory;
  2. sends that audio to a speech engine (whisper.cpp), which is included with talkflow and runs as a separate program on your own computer, at the address 127.0.0.1 (also called "localhost"), which always means your own computer;
  3. turns the speech into text using OpenAI's Whisper speech model, running on your computer: the English-only "small.en" version and, on computers fast enough for it, the larger "large-v3-turbo" version for the final text;
  4. types the text into the app you're using.

Audio is never saved to disk. It stays in memory only, until your next dictation or until you quit talkflow.

The speech engine is a program on your computer. It is not a cloud server, and your audio doesn't travel over the internet, unless you turn on OpenAI transcription with your own key (see Your own OpenAI or Anthropic key).

No voiceprints. talkflow uses your voice only to turn it into text. It doesn't create a voiceprint or any other biometric identifier, and it doesn't try to recognize who is speaking.

What we receive from the app

One thing: the first time a newly installed copy of talkflow runs while online, it tells us once, so we can count how many times talkflow has been installed. This applies to versions 0.1.6 and later; earlier versions send us nothing.

Apart from that, the app doesn't send your audio, your text, your stats or any other information to us. There are no accounts, no usage analytics and no crash reports sent to us. Besides the speech engine on your own computer and this one-time request, the app connects only to GitHub to check for updates (below), and to OpenAI or Anthropic if you add your own API key in Settings.

Update checks

When it is online, the app asks GitHub whether a newer version exists: shortly after it starts, every few hours while it runs, and when you open the dashboard. It downloads a small public file listing the latest release (if that file is missing, it asks GitHub's release API instead). The request sends nothing about you or your dictation. Like any web request, it reaches GitHub with your IP address, and it names the app and its version (for example "talkflow/0.1.5"). GitHub's privacy statement covers that request; we don't receive it.

When a newer version exists, the app says so. Nothing is downloaded or installed until you click to update, and the download is checked against the release's published checksum before it is installed. Without an internet connection, the app works the same and simply checks again later.

Apart from the one-time install count, which can't be linked to you, we don't receive anything from the app, so we can't see, recover or delete anything on your computer for you. You are in control of it.

Your own OpenAI or Anthropic key

In Settings you can add your own API key for OpenAI or Anthropic and turn on a feature that uses it. These features are off by default. When one is on, talkflow sends data directly from your computer to that company, using your key. It doesn't pass through us, and we don't receive it.

Each request carries your key and the technical details any web request has, such as your IP address. When you save a key, talkflow checks it by asking the company for its list of models, which sends no dictation.

What OpenAI or Anthropic does with that data is governed by your own agreement with them, under the account your key belongs to, and by their policies, including how long they keep it. See the OpenAI privacy policy and the Anthropic privacy policy. We don't control them, and we aren't responsible for how they handle it. Don't turn these features on for dictation you aren't allowed to share with them, for example confidential work or health information. If a request fails, talkflow transcribes or punctuates on your computer instead.

Setting up the app

The whisper.cpp speech engine comes inside the talkflow download, so nothing else needs to be installed. The Whisper speech models are downloaded from Hugging Face: small.en (English only) during setup, and large-v3-turbo in the background on computers that use it. Hugging Face sees your IP address when you download them and handles it under the Hugging Face privacy policy.

What stays on your computer

The app keeps a small amount of information on your own computer. On a Mac, your data is in ~/Library/Application Support/talkflow/. On Windows, it is in %APPDATA%\talkflow\, and the models and logs are in %LOCALAPPDATA%\talkflow\.

talkflow doesn't encrypt these files itself. They are protected by your user account, and by disk encryption (FileVault on a Mac, BitLocker or device encryption on Windows) if you have it turned on.

Deleting them. You can delete these files at any time. On a Mac, deleting the talkflow app may not delete them. On Windows, uninstalling removes the models and logs but keeps your data folder, so remove %APPDATA%\talkflow\ too if you want it gone.

What the app reads while you dictate

When you start a dictation, talkflow reads what is around your cursor, through macOS Accessibility on a Mac or UI Automation on Windows: up to 300 characters before the cursor, up to the last 4,000 characters of the text field, and names visible on screen, such as an email's recipients. It uses this to get spacing and capitals right, to spell names correctly and, a little later, to notice when you corrected a word it got wrong, so it can learn that word. It never reads password fields, and it never reads whole documents beyond those limits. This happens in memory. It isn't logged or sent anywhere, and it is kept only for that dictation, except a corrected word, which is saved to your learned words.

The clipboard. On a Mac, talkflow types with keystrokes and doesn't use the clipboard, with one exception: if the text can't be typed (for example, because you switched to another app while speaking), it is put on the clipboard so your words aren't lost. On Windows, longer dictations are pasted: talkflow saves what is on your clipboard, puts the dictation there, pastes it, and then puts your clipboard back. That paste is kept out of Windows clipboard history and the cloud clipboard.

Permissions the app asks for

On a Mac. You can turn either of these off at any time in System Settings > Privacy & Security, but talkflow won't work without them.

On Windows. Windows doesn't ask desktop apps for permissions. Instead:

Things outside talkflow's control

Some things that happen on your computer are up to you, your organization or other companies, not us:

Future versions

This part describes the app as of the "Last updated" date above. The one-time install count starts with version 0.1.6. If we ever add a feature that sends any other information off your computer, such as crash reports, we will update this policy before releasing that version and describe the change in the release notes.


Emails you send us

If you email us, we receive your email address, your name if you include it, and whatever you write. We use it only to reply and to deal with your request. Our legal basis is our legitimate interest in answering you.

Our address forwards to a Gmail inbox. Google, and the email service that forwards the messages, process them on our behalf. We keep emails only as long as needed to handle your request.

Selling and sharing

We don't sell your personal information. We don't "share" it for cross-context behavioral advertising, as California law defines those terms, and we don't use it for targeted advertising. We haven't done any of these things in the past 12 months.

How long we keep information

We keep personal information only as long as we need it for the purposes described in this policy, including to meet legal obligations, resolve disputes and protect the site. To decide how long that is, we consider how much information there is, how sensitive it is, the risk of harm from misuse, and whether we can achieve our purpose another way. When we no longer need it, we delete it or make it anonymous. The specific periods for website logs and emails are described above.

Sensitive information

Please don't send us sensitive personal information by email, such as health, financial or government ID details, or information about your race, religion, political views or sex life. We don't need it to answer you.

No automated decisions or profiling

We don't make decisions about you by automated means, and we don't profile you, in ways that have legal or similarly significant effects on you.

Children

The site and the app are not aimed at children under 13, and we don't knowingly collect personal information from children under 13. If you believe a child under 13 has sent us personal information, for example by email, contact us and we'll delete it.

Your rights

Depending on where you live, you may have the right to:

In practice we hold very little. We have no app data about you at all (the install count is a single total that can't be linked to anyone), and we don't keep copies of website request logs. If you contact us about website logs, we may not be able to find them, because they are held by our hosting provider.

How to make a request. Email samir@talkflow.live. We'll reply within one month. If we need to confirm who you are, we'll ask only for what we need. Where the law allows, someone you authorize can make a request for you; we may ask for proof that you authorized them. If we turn down a request, we'll tell you why. Where the law gives you the right to appeal that decision, reply to our answer and say you want to appeal, and we'll review it again.

Complaints. If you have a concern about how we handle your information, please tell us first at samir@talkflow.live. You also have the right to complain to a data protection authority:

If you live in California, Texas or another US state with a privacy law, you can make the same requests by email, including an appeal. We'll handle them in the same way, whether or not that law applies to us. If we deny your appeal, you may contact your state's attorney general.

Keeping information safe

No system is perfectly secure, and we can't guarantee the security of information sent over the internet. We keep the information we hold to a minimum, which is itself a protection.

International transfers

Vercel is based in the United States and serves the site from data centers around the world, so website request information may be processed outside your country, including in the United States. Vercel says it is certified under the EU-U.S. Data Privacy Framework (and its UK and Swiss extensions) and also uses standard contractual clauses approved by the European Commission. Emails you send us are handled by Google, which also processes data in the United States and other countries. You can ask us for more information about these safeguards at samir@talkflow.live.

Changes to this policy

We may update this policy. When we do, we'll post the new version on this page and change the "Last updated" date at the top. Significant changes to the app's data handling are also noted in the release notes.

The app tells you when a new version is available, but not when this policy changes. Please check this page or the release notes when you update. See also our Terms of Use.

Contact

Samir Kharel
Email: samir@talkflow.live