Privacy Policy

2.1Blockchain information

• Wallet address
• Public blockchain transactions
• NFTs and tokens you hold
• History of interactions with our smart contracts

2.2Usage information

• IP address (anonymized)
• Browser and device type
• Pages visited and time spent
• Language and theme preferences

2.3Cookies and local storage

• Cookie consent preferences
• Authentication session (temporary)
• User interface settings
• Data cache for better performance

4.1What is collected

• The path of the page visited, without parameters (anything after “?” and “#”) and with sensitive parts replaced by “:redacted” (wallet addresses, transaction hashes, .eth names, email addresses, phone numbers, codes and tokens)
• The referring page (referrer) reported by your browser, only on the first page of a visit and only when you arrive from another site; it may include the address of the previous page. On our own pages, sensitive parameters (such as the payment return or signing challenges) are removed from the address bar as soon as they are read
• Country, device type, browser and operating system. According to Vercel, the IP address is not stored and visitor counting uses an identifier discarded within 24 hours
• Page performance metrics (Core Web Vitals)
• Usage steps: edition viewed, purchase started/completed/declined, wallet connected, Live unlocked, first message, voice played and agent named — with only the edition, payment method, reason code, channel (text or voice) or voice engine. The Founder edition, the smallest, is recorded only as “high”, except when viewing the presale

4.2What is never collected

• Wallet address, transaction hash, email address, agent name, message text or purchase amount
• Nothing from the admin area, the API, the vault, the secure signer or internal tools; the payment return page does not count a visit
• Nothing when your browser sends Do Not Track or Global Privacy Control (the scripts are not even loaded)
• This data is not used for advertising and is not combined with Google Analytics or the pixels

6.1Feedback on replies (training labels)

• Under each AGEN reply you can choose thumbs up, thumbs down or "Not what I meant". These are kept as training labels only if you turn on "Help train AGEN" in Settings. The option is off by default; while it is off, your browser does not send these labels.
• A label contains only the type of feedback, an optional reason from a fixed list, the id of the reply, in some cases the ids of memories the reply used, and the time. No text from your conversations is stored as a label or used for training.
• Labels are encrypted on our server and linked to your NFT.
• Labels are kept for at most 12 months and are then deleted automatically.
• Turning the option off deletes all your stored labels immediately. You can also delete them at any time with "Delete my feedback".
• We keep a record of when the option was turned on or off, and by which wallet, as proof of your choice.
• When the NFT is transferred to a new owner, its labels and this consent record are deleted. The new owner starts with the option off.
• Labels are only used in totals combined across at least 20 owners who turned the option on, never on their own.
• "Pin as fact" is not training data: the fact you write is saved in your agent's private memory, works whether the option is on or off, and can be removed in Memory.

6.2Encrypted memory archive

• AGEN can keep an archive of your conversations so your agent remembers earlier exchanges. This archive is currently switched off; the points below describe how it works when it is on.
• The archive is encrypted separately for each NFT. Encryption protects the data on disk and in backups; our server can decrypt it to answer you, so it is not end-to-end encryption.
• There are two retention modes: in the 30-day mode, conversations older than 30 days are deleted automatically together with the summaries made from them; in the "keep all" mode nothing expires and only you delete.
• As the owner you can delete single items, a day, week, month or year, or everything. Summaries made from deleted items are rebuilt without them.
• Deleting removes the data from the archive on our server. It does not reach copies kept in the logs of the AI model providers that generate replies, or in our backups; those copies are removed only when the provider's or the backup's own retention period ends.

6.3"Share usage data" setting

• The "Share usage data" setting in Settings is off by default. If you saved your settings before October 12, 2026, when it was shown on by default, check its value in Settings.
• Today it only records your choice with your agent's settings; no additional usage data is collected or sent because of it.
• It is separate from cookie consent (section 3) and from "Help train AGEN" above.

6.4Voice: microphone, speech recognition and the agent's voice

• Voice is optional. The microphone is only used when you press the microphone button or tap the orb to talk live, and your browser asks for permission first. While live, a "Microphone on" indicator stays visible; listening stops when you mute it, hide the tab or leave live mode. Unless you turn on "I'm on headphones", listening also stops while your agent is speaking.
• Speech-to-text is done by your browser, not by QuantumStack. Where the browser supports on-device recognition (Chrome 139 or later, with the language pack for your language installed), it runs on your device and your microphone audio does not leave it. Otherwise your browser's speech provider does it and receives your microphone audio under its own privacy terms (for example Google in Chrome, Microsoft in Edge, Apple in Safari). The notice before your first voice use tells you which of the two applies. Firefox has no speech recognition, so live mode there only accepts typed messages.
• QuantumStack does not receive or store your microphone audio. The site only receives the text your browser transcribed, which is sent to your agent like a typed message.
• Your agent's spoken replies are generated by ElevenLabs: the text of each reply sentence is sent from our server to ElevenLabs, and the audio is streamed back to your browser. The site does not store this audio; our server logs only technical data (engine, status, length), never the text or the audio.
• When the server voice is unavailable, your browser's own voice may read the reply instead; some of these voices are online services of the browser maker.

8.1AI models and voice

• Anthropic (United States), when active: generates replies. It receives the full chat turn: your message, your agent's persona, the facts you gave your agent and the memory relevant to the reply. Recognized secret patterns (such as passwords, keys and tokens) are redacted before sending.
• OpenAI (United States), when active: receives the same full chat turn when you choose the premium model (GPT-4o), which is opt-in.
• Google Gemini API (United States / global): our main model for everyday replies and for voice. It receives the full chat turn. We currently use Google's free tier: under Google's terms, content sent on this tier and the responses may be used by Google to improve its products and may be read by human reviewers. We plan to move to the paid tier, where this does not apply.
• A multi-model security feature, only when enabled, may split long messages across Anthropic, OpenAI, Google and DeepSeek; each of them then receives part of the message.
• ElevenLabs (United States): generates your agent's voice. It receives the text of the reply to be spoken, never your microphone audio. ElevenLabs keeps data under its own retention terms.
• Older voice route (not used by the AGEN Live microphone or voice call): it uses Google Cloud Speech-to-Text and Google Cloud Text-to-Speech when configured, and Replicate (United States; its openai/gpt-4o-transcribe model forwards the audio to OpenAI) only if enabled. In AGEN Live, if ElevenLabs fails, your browser's own voice reads the reply instead.
• Your browser's speech provider, when recognition is not done on your device (for example Google in Chrome, Microsoft in Edge, Apple in Safari): it receives your microphone audio under its own terms.
• DeepSeek (China): currently used as a backup model for replies when Google Gemini fails, hits its free-tier limit or is slow to respond (for voice, after a few seconds); in that case both may receive the same turn. It receives the full chat turn. Data is processed in China. We plan to stop using it for personal data as soon as another provider is funded.

8.2Hosting, storage, blockchain and payments

• Vercel (United States / global network): hosts the site and its API routes, and provides the cookieless measurement described above.
• Upstash (Vercel KV): limits the rate of anonymous requests. It stores short-lived counters keyed by a hash of your IP address, never the IP address itself.
• Hostinger (EU): the virtual private server that runs the AGEN agent. Your agent's conversations and memory are processed and stored there, encrypted at rest with keys held by our server.
• Supabase: stores purchase, presale and profile records, activity metadata without the content of your conversations, and the email address you give us for the launch list. The storage region follows our hosting configuration.
• Pinata / IPFS: IPFS is a public network, and anyone with a file's address can fetch it. Backups of your agent's profile are encrypted before upload, so only ciphertext goes there. The NFT metadata and image you choose are public and should be considered permanent: once published, copies can remain on the IPFS network.
• Blockchain node providers (RPC): to read the blockchain and send transactions, the site and our servers connect to Ethereum node providers, which receive the IP address of the device making the request and the wallet addresses involved. The site uses Infura, PublicNode, dRPC, 1RPC and LlamaRPC, and our server uses Alchemy for minting; the AGEN server uses Infura, PublicNode and dRPC.
• Wallet connection: if you connect through WalletConnect (Reown) or Coinbase Wallet, those services receive your wallet address, connection data and IP address under their own terms. MetaMask connects through its own browser extension or app, under MetaMask's terms.
• Stripe: processes card payments. For payment data (such as your card details) Stripe acts as an independent controller under its own privacy policy; we never receive your full card number.
• Ethereum network: wallet addresses, transactions and NFT ownership are public on the blockchain, permanent and cannot be deleted by us or by anyone (see “Important: blockchain data”).
• Google (Analytics, Ads), Meta and TikTok: only if you consent, as described in “Analytics and marketing cookies”.

8.3International transfers

Several of these providers are in the United States or China or operate globally, so your data may be transferred outside your country, including outside the European Economic Area and Brazil, to the United States and China. Where applicable, these transfers rely on the providers' data processing agreements and on the European Commission's standard contractual clauses or equivalent safeguards provided for by law (for example, under the LGPD). Providers that act independently, such as your browser's speech provider, Stripe, Meta and TikTok, transfer data under their own terms. For DeepSeek and for Google's free Gemini tier we have no data processing agreement and no standard contractual clauses, and China has no adequacy decision from the European Commission or from Brazil's ANPD.