Duck.ai is DuckDuckGo’s chat interface to models it does not build itself — it sits between you and Anthropic, OpenAI, Mistral, Azure and Tinfoil, forwarding your prompt with the identifying parts taken off. It launched in autumn 2024 and left beta in March 2025. The design question it answers is narrow and specific: what would a chat client look like if the thing you were trying to hide was not the content of the conversation but the fact that it was yours? The answer here is no account, no IP address passed downstream, and no chat history on anyone’s server.

Privacy Architecture

Nothing is asked of you before the first message. No email, no phone number, no card — you open the page and type. That matters more than it sounds, because an account is the durable identifier that ties every session you ever have back to one profile, and without one there is nothing for a chat history to accumulate against.

What reaches the model provider is the prompt and not much else. DuckDuckGo’s documentation states that metadata containing personal information — the IP address is the example given — is completely removed before the model is prompted, so the provider sees requests arriving from DuckDuckGo rather than from you. The company’s own privacy policy goes further and says IP addresses are never logged to disk alongside Duck.ai chats in a way that could be tied back to a user. So the proxy step is not just a relabelling exercise where the operator keeps the mapping — the stated position is that the mapping is not written down at either end.

Conversations stay on the device. Recent Chats live in local browser storage, not on DuckDuckGo servers; you can delete them one at a time, burn all of them with the Fire Button, and pin up to five you want to keep. The practical consequence is the one people notice second: there is no sync. Start a conversation on a laptop and it does not exist on the phone. That is the direct cost of not having a server-side account, and it is not a bug in the design so much as the design.

On the provider side, the contractual position is zero data retention — prompts and responses deleted immediately after a reply is generated — with one documented exception. Prompts may be held in short-term memory for up to an hour for caching, to cut token costs and latency, and DuckDuckGo states this cache is never written to disk. A one-hour window in volatile memory is a real exposure and it is better to see it written down than discovered later. The no-training commitment covers all five providers: Anthropic, OpenAI, Mistral, Azure and Tinfoil are all contractually barred from using Duck.ai conversations as training data.

The hardware-enforced part of this needs to be stated precisely, because it applies to two models and not to the other eight. Requests routed to Tinfoil — which hosts gpt-oss-120b and Gemma 4 31B — are processed inside a Trusted Execution Environment, so that Tinfoil itself cannot see the prompt or the model’s response. That is a stronger architectural claim than a contract alone, though it still rests on DuckDuckGo and Tinfoil’s own description of the setup rather than an independent audit of it. For the Claude, GPT and Mistral models, which run through the Anthropic, OpenAI, Mistral and Azure APIs, there is no such enclave: the protection there is the contract plus the stripped metadata, which is a weaker thing. If enclave-level isolation is what you came for, the choice of model determines whether you get it.

Trust & Transparency

The operator is Duck Duck Go, Inc., a United States company. That is the jurisdiction people will weigh first, and it is worth weighing against what the architecture actually leaves available to compel: no account, no server-side chat store, and no IP-to-query mapping written to disk. A demand for records still runs into the question of which records exist.

Ownership is unusual for a company of this age and size. DuckDuckGo has been independent since it was founded in 2008 and remains majority-owned by its founder and team members — no acquisition, no private-equity parent, no holding company whose other properties sell advertising data. There is no telemetry on how the chat is used, and the company states plainly that it does not track usage, store prompts, or train on the data.

The documentation is the strongest part of the trust case. The privacy pages name each model provider, state the retention terms per provider, disclose the one-hour caching exception rather than rounding it down to “we don’t store anything,” and scope the enclave claim to the specific host it applies to. That is more granular than this category usually gets.

What is missing is verification. No independent third party has audited Duck.ai’s data flow, and the implementation is closed source, so the metadata-stripping step and the no-disk-logging claim cannot be checked from outside. This is not a failing unique to this product — no AI chat assistant in this category currently has an independent audit, and honesty requires saying that the whole vertical is in the same position rather than singling out one operator. But it does mean that every architectural claim above rests on the operator’s own description of its systems, backed by a reputation built in another product line, and not on anyone’s inspection. That is the reason this score sits where it does rather than higher.

Features

Ten models are available across the tiers, from four different families. Six come free: Claude 4.5 Haiku, Mistral Small 4, GPT-5.4 nano, GPT-5.4 mini, and the two Tinfoil-hosted open-weight models, gpt-oss-120b and Gemma 4 31B. The subscription adds GPT-5.6 Terra and Claude Sonnet 4.6, and the higher tier adds Claude Opus 4.8 and GPT-5.6 Sol along with higher reasoning effort.

Switching between them is free-form — you pick a model per conversation rather than committing to one, which is the practical argument for a front-end like this over going to each provider directly. Comparing how two families answer the same question takes one dropdown instead of two accounts.

File uploads are supported, with limits worth knowing before you rely on them. Images and PDFs go through, but only on some models rather than all of them, and PDF page counts are capped by tier: 15 pages free, 35 on Plus, 50 on Pro. Spreadsheets are not supported at all. For document analysis at any volume, those page ceilings arrive quickly.

Image generation is available, also with an unpublished daily cap. There is no confirmed live web search inside the chat interface itself — DuckDuckGo’s AI-assisted answers on the search results page are a separate feature, and nothing in the Duck.ai documentation establishes that the chat mixes fresh search results into its replies, so plan on the models’ training data being the boundary.

Usability

There is no Duck.ai application to install. Not on Windows, macOS or Linux, not on iOS or Android. The product is a web page, plus an entry point built into DuckDuckGo’s own browser for people already using it. If your habit is a dock icon, a keyboard shortcut and a native window, that habit does not transfer here.

Set against that, the path to a first message is about as short as it gets: open the site, pick a model, type. No signup wall, no email confirmation, no onboarding sequence, no card on file to try the paid models’ free counterparts. For the case of “I have one question and I do not want an account for it,” the absence of an app costs nothing and the absence of a signup saves everything.

The device-local history has a usability consequence beyond the missing sync. Clearing browser data clears your chats, the Fire Button clears them by design, and a private window starts empty every time. Five pinned conversations survive within a browser profile, which is enough to keep a working reference thread and not enough to treat this as a place where research accumulates.

Daily usage limits exist on the free tier and DuckDuckGo does not publish the numbers — only that a cap is there, that the paid tiers raise it, and that the top tier doubles the middle one. You find out where the ceiling is by hitting it.

Price & Value

The pricing here is easy to misread, so state it plainly: $9.99 a month is not the price of Duck.ai. It is the price of DuckDuckGo’s entire subscription, a four-in-one package covering a VPN, advanced Duck.ai models, a personal information removal service, and identity theft restoration. Access to the better models is one quarter of that bundle. The Pro tier at $19.99 a month is the same four services with Claude Opus 4.8, GPT-5.6 Sol, higher reasoning effort and double the usage limits on the chat side. Annual billing is $99.99 and $199.99 respectively.

There is no AI-only plan. If the VPN, the data removal service and the identity restoration service are things you neither want nor will use, there is no way to pay a smaller amount for just the models — you buy all four or none. Evaluated as an AI subscription in isolation, that is poor value; evaluated as a privacy bundle where the chat models come along with three other services you were going to consider anyway, the arithmetic changes completely. Which of those two situations you are in decides whether this price is reasonable.

The free tier is the part that needs no qualification. Six models, including two running in hardware enclaves, usable with no account, no card and no trial clock — the only limit is an unpublished daily message cap and image-generation cap. For a large share of everyday use, the free tier is the product, and the subscription buys frontier models rather than buying access at all.

Pros and Cons

No account of any kind — no email, phone number or card before the first message, so there is no durable identifier to accumulate history against
IP address and other personal metadata are removed before the prompt reaches the model provider, and the privacy policy states IP addresses are never logged to disk in a form tied back to a user
Chats are stored in the browser on your own device rather than on a server, deletable individually or all at once with the Fire Button
All five model providers — Anthropic, OpenAI, Mistral, Azure and Tinfoil — are contractually barred from training on Duck.ai conversations
gpt-oss-120b and Gemma 4 31B run inside a Trusted Execution Environment at Tinfoil, so the host cannot see the prompt or the response
Six models free with no account, including both enclave-hosted ones, with no trial period to expire
Ten models across four families with per-conversation switching, so comparing two answers takes one dropdown rather than two accounts
Independent since 2008 and majority-owned by its founder and team, with no advertising-data parent company and no usage telemetry in the chat

No independent third-party audit of the data flow exists, and the implementation is closed source, so every architectural claim rests on the operator's own description — a gap shared across this whole category, but a gap
The Trusted Execution Environment covers only the two Tinfoil-hosted models; the Claude, GPT and Mistral models are protected by contract and metadata stripping alone
No native desktop or mobile application for Duck.ai — it is a web interface, plus an entry point inside the company's own browser
Device-local storage means no sync between machines: a conversation started on a laptop does not exist on the phone, and clearing browser data erases it
Advanced models are only sold inside a $9.99/month four-service bundle with a VPN, personal information removal and identity theft restoration — there is no cheaper AI-only tier
Prompts may sit in provider-side cache memory for up to an hour, which is disclosed but is still a window where the text exists outside your device
Daily usage limits on the free tier are real but the numbers are not published, so you discover the ceiling by reaching it
PDF uploads are capped at 15 pages free, 35 on Plus and 50 on Pro, only some models accept files at all, and spreadsheets are not supported
No confirmed live web search inside the chat, so answers are bounded by each model's training data
United States jurisdiction, which is the weakest part of the picture on paper — mitigated in practice by how little data exists to compel

Our Rating

Duck.ai suits the person who wants to ask a model a question without that question becoming part of a profile: no account, nothing stored on a server, identifying metadata removed before the prompt leaves, and a free tier good enough that most people never reach the paywall. The two enclave-hosted models push that further than contractual promises alone can. It suits you less if you want conversations synced across devices, a native app, an AI subscription priced on its own rather than bundled with three unrelated services, or an independent auditor’s word rather than the operator’s — that last one is the honest limit here, and it applies to every product in this category rather than to this one alone.

Privacy Architecture8.5/10
No account, IP and other identifying metadata stripped before the prompt reaches the model provider, chats kept on the device rather than on a server, and contractual no-training terms with every provider — with a hardware-enforced enclave on two of the ten models rather than all of them
Trust & Transparency6.2/10
Detailed public documentation of the data flow, a no-telemetry stance and long-standing independent ownership, but the implementation is closed source and no third-party audit of these claims exists, so the architecture has to be taken on the operator's word
Features7.0/10
Ten models from four providers with free switching between them, image and PDF uploads, and image generation — but no spreadsheet support, page-count caps on PDFs and no confirmed live web search in the chat interface
Usability6.5/10
Opening a chat takes no signup and no configuration, but there is no native desktop or mobile app for Duck.ai itself — it is a web interface, plus an entry point inside the company's own browser
Price & Value8.0/10
Six models usable free with no account and no card, but the advanced models are only sold inside a $9.99/month four-service subscription with no cheaper AI-only option

Overall7.3/10

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

See our rating methodology →