Obscurify.ai is a chat service on the web and on Android that stands between you and the companies that run AI models. Your message goes to Obscurify first. Obscurify removes your IP address and sends the request on under its own shared credentials, so the model provider sees a request from Obscurify, not from you. In that sense it works as a proxy for AI chat. The service is run by CryptoTalks LLC, a company registered in Washington, DC, and launched in November 2023 under the name CryptoTalks.

The privacy case rests on three things. You can use the service without giving any identity, you can pay in a way that the operator says it cannot link to your account, and your chats are not saved on its server. The third point is the operator’s statement: the server code is closed, and by Obscurify’s own account the service has no hardware isolation and no end-to-end encryption, so the prompt is readable on its server while it is processed.

The catalogue is large: 296 models in the picker, from GPT, Claude and Gemini models to uncensored community fine-tunes, plus a Local Mode that runs open models inside the browser. Guests chat for free with one open model. Paid use starts at $0 a month with per-message billing from a prepaid balance; the Casual and Pro plans cost $20 and $45 a month.

Privacy Architecture

There are two ways in, and neither asks for an email address. Guests open the site and type, with no account and no cookies, limited to one open model (Llama 3.3 70B) and 30 messages an hour. An account is a random 16-digit number, with no username, password or email; the server keeps only a scrambled copy of that number, so it cannot read it back. Older accounts may still carry a username, a hashed password and an optional email from before, and Obscurify is retiring password sign-in. Since an account is optional, you can try the service before giving it any lasting identifier.

Every request to a hosted model passes through Obscurify. The operator says it strips your IP address and other identifying details before forwarding, and the provider receives the request under Obscurify’s credentials, so it sees Obscurify’s request rather than yours (see IP anonymization). Obscurify itself is a different matter. The operator’s own comparison page says the service has no hardware-isolated environment and no end-to-end encryption, so there is no hardware attestation to check. In practice the operator can read your prompt while it passes through. You are trusting the company’s description of what it does with that text, and nothing technical backs that trust.

Chats are not stored on the server. On the web, history is optional and lives in your browser’s local storage; in the Android app, chats stay on the phone. In this storage model, chat history exists only on your device, apart from the 30-day IP and user-agent logs described below.

Some data is logged. The privacy policy says the IP address and user agent go into security and error logs for 30 days to prevent abuse. Guest chat holds the IP address in memory for up to an hour to enforce the 30-message limit. At sign-up, a scrambled value derived from the IP address sits in memory for 24 hours so the same person cannot claim the starter credit again and again.

What happens at the model provider is set by the provider. Obscurify says it routes text requests only through zero-data-retention routes, where the provider’s published policy promises not to train on or keep the data, and that models without such a route, including many free ones, are not offered. It has no separate contracts with the providers. Whether a provider keeps or trains on your prompt therefore depends on that provider following its own published policy, so there is no contractual commitment that providers do not train on prompts. Providers may also keep parts of a prompt in short-term memory to avoid repeating work; for Claude models that is five minutes, and Obscurify does not attach your account to those requests. Image requests follow the image provider’s own policies. Stripping the IP address does not touch the text itself, so a name or address you type into a prompt still reaches the provider and may come back in a reply.

Payment is kept apart from the account through blind vouchers. You buy a voucher, Obscurify signs it in a hidden form it cannot recognise later, and you then redeem it into your account. The operator sees the payment, the IP address used for the purchase and the time of redemption, but says it cannot tell which payment funded which account. To break the link between purchase time and redemption time, the Android app redeems each voucher at a random moment within the hour. Vouchers can be bought with Monero, Bitcoin or Lightning through Obscurify’s own payment server, or by card through Stripe, which receives your card details.

Two options reduce what Obscurify itself sees. The whole site is available as a Tor onion service, and a separate page works without JavaScript for Tor Browser at its safest setting. Local Mode goes the other way: it downloads an open model and runs it inside the browser, and nothing is sent to the server. Its speed and the size of model you can run depend on your own device.

Trust & Transparency

The code that runs on your side is public. The web client, meaning the part that runs in your browser, is published under the GPL-3.0 licence, and its README says updates are synced from the code running on obscurify.ai. The Android app is a native Rust client, also under GPL-3.0, and its releases are signed with a published certificate. The server, where IP stripping and in-memory processing happen, is closed. That makes the code only partly open source, and the claims that matter most cannot be checked from outside.

No independent security audit of Obscurify has been published. There is also no public record of a data breach or security incident.

The privacy policy names its own limits: the 30-day logs, the missing provider contracts, the cache lifetimes and the fact that typed personal information reaches the provider. There are no tracking cookies, third-party analytics or ad trackers, so telemetry is limited to coarse counts: the request type, whether it came over Tor and, for page visits, the referring domain, with no account, IP address or cookie attached.

Obscurify publishes a warrant canary signed with PGP. The current one is dated 1 October 2026 and says the company has received no National Security Letters, FISA orders, gag orders, government data requests, backdoor requests or subpoenas; the next version is due by 1 November 2026. The signing key is published by the operator, so a valid signature shows the canary came from the same key holder each time, not who that holder is.

Who runs the service remains unclear. The about page names CryptoTalks LLC and its Washington, DC address but no founders, staff or owners. The company is in the United States, and the terms of service name no governing law, so a user has no stated legal framework to rely on in a dispute.

Features

The model picker lists 296 models. Eight are marked Frontier and thirteen Uncensored. The catalogue spans the GPT, Claude, Gemini, Llama, DeepSeek, Mistral and Qwen families and many community fine-tunes, from about 60 model providers in all. You can switch models inside the same chat. Access depends on the plan: Frontier models need Casual or higher, and Uncensored models need Pro.

Web search is a tool in the chat, so a model can pull in current pages instead of answering only from its training data. You can attach images, but there is no general file upload: PDFs, spreadsheets and other documents cannot be sent to a model.

Local Mode gives a chat with an open model that never leaves the browser. It is the option for questions that even the operator should not see.

Every plan with a positive balance gets an OpenAI-compatible API at the same per-model prices as the web chat.

Usability

The web chat works in any browser, and the onion address and the no-JavaScript page cover Tor Browser users. The only app is for Android. It is a file of about 5 MB that you download from GitHub and install by hand, needs Android 7 or newer, and is not on Google Play; a listing in the F-Droid app store is described as coming. The app does not update itself, so Obscurify suggests the Obtainium app, which fetches new releases for you. In answers, the Android app shows remote images as links so that loading them does not reveal your IP address. There is no iOS app and no desktop app.

History has the usual cost of local storage. A chat started in the browser on a laptop is not on the phone, and clearing browser data erases the web history.

The account number is both login and recovery. Obscurify stores only a scrambled copy of the number and cannot recover it, so a lost number means a lost account, along with any balance on it.

Guests can test the service without signing up, but 30 messages an hour with one model is enough for a trial, not for regular work. Paid plans move requests up the queue: Casual gets a faster queue and Pro the highest priority.

Price & Value

Pricing has four levels. Guests pay nothing. Pay As You Go costs $0 a month: each message is charged at its model’s price from a prepaid balance, and API access is included. Casual costs $20 a month, includes $15 of usage and unlocks Frontier models. Pro costs $45 a month, includes $35 of usage and unlocks Uncensored models. Usage beyond the included amount is drawn from the prepaid balance. New accounts get $1 of starter credit while the monthly allowance for it lasts.

Plans renew monthly from the balance, with no annual price and no introductory rate, so the price you see is the price you keep paying. On Casual, $15 of the $20 comes back as usage credit, so the plan itself costs $5 for Frontier access and a faster queue; on Pro, $35 of the $45 comes back, leaving $10 for Uncensored access and top priority. Someone who mostly uses open models can stay on Pay As You Go and pay only for the messages sent.

Money is hard to get back. Deposits are non-refundable and expire five years after they are made, and a lost voucher is not replaced. With a lost account number, the balance is gone as well.

Pros and Cons

Guests can chat without an account, email or cookies, and accounts are a random 16-digit number with no email or password
Obscurify removes your IP address and forwards requests under its own credentials, so model providers do not see who sent them
Chats are not saved on the server; history stays in the browser or on the phone
Blind vouchers bought with Monero, Bitcoin, Lightning or card keep the payment apart from the account, by the operator's description
The whole site is available as a Tor onion service, with a no-JavaScript page for Tor Browser's safest setting
Local Mode runs open models in the browser with nothing sent to the server
296 models in one picker, including Frontier and Uncensored ones, with switching inside the same chat
Web search and image attachments are built into the chat
An OpenAI-compatible API comes with every plan that has a positive balance, at the web chat's prices
The web client and the Android app are public under GPL-3.0
The privacy policy states its own limits, and a PGP-signed warrant canary is published with a renewal date

No hardware isolation and no end-to-end encryption, so the operator can read prompts while they pass through its server
The server code is closed, so IP stripping and in-memory processing cannot be checked from outside
No separate contracts with model providers, so whether they keep or train on prompts depends on their own published policies
No independent security audit has been published
No founders, staff or owners are named, and the terms name no governing law
IP address and user agent are kept in logs for 30 days
There is no general file upload, so documents such as PDFs cannot be sent to a model
The only app is an Android APK from GitHub with no automatic updates, and there is no iOS or desktop app
A lost account number cannot be recovered, and the balance on it is lost too
Frontier models need the $20 Casual plan, and Uncensored models need the $45 Pro plan
Deposits are non-refundable and expire after five years

Our Rating

Obscurify.ai suits someone who wants to use GPT, Claude or Gemini models without an email address, a card in their name or a chat history on someone else’s server, and who is willing to pay in cryptocurrency or reach the site over Tor. Its wide model range, web search and API make it usable for daily work, and Local Mode covers questions that should not leave the device at all. It does not suit someone who needs the operator to be technically unable to read prompts, because nothing here prevents that, or someone who wants to know who runs the service. It also does not suit iPhone or desktop-app users, anyone who expects chats to follow them across devices, or anyone who needs to send documents to a model.

Privacy Architecture7.5/10
Obscurify asks for no identity, removes the IP address before a request reaches the model provider, keeps no chats on its server and separates payment from the account with blind vouchers, but the operator can read prompts while processing them and the no-training promise rests on the providers' own policies rather than a contract.
Trust & Transparency5.8/10
The web and Android clients are public under GPL-3.0, the privacy policy states its own limits and a PGP-signed warrant canary carries a date by which it will be renewed, but the server code is closed, no independent audit exists, nobody behind the company is named and the terms name no governing law.
Features8.0/10
The picker holds 296 models with switching inside one chat, plus web search, image attachments, an in-browser Local Mode and an OpenAI-compatible API on any positive balance, while the main gap is the lack of a general file upload.
Usability6.0/10
The web chat works in any browser and over a Tor onion address, but the only app is an Android APK installed from GitHub without automatic updates, history does not move between devices, and a lost account number takes the balance with it.
Price & Value6.0/10
Guests chat for free and Pay As You Go costs nothing beyond the messages used, but frontier models need the $20 Casual plan, uncensored models need the $45 Pro plan, and deposits are non-refundable.

Overall6.8/10

Privacy Architecture 30% · Trust & Transparency 20% · Features 20% · Usability 20% · Price & Value 10%

See our rating methodology →