
Fluid
Fluid is a private AI assistant for Apple Silicon Macs that runs its language model entirely on the device. A global Ctrl+Space shortcut opens it anywhere, dictation stays local, and a paid EU-hosted API adds a larger model.
What is Fluid?
Fluid is a desktop AI assistant for macOS whose defining choice is where the computation happens: on your own machine. The homepage sums it up as a private AI assistant for Mac, 100% offline. The app ships with a Meta Llama 3 model — Llama 3.1 from version 0.2.6 — downloads it on first launch, and answers from there. Pressing Ctrl+Space opens it over any application, Spotlight-style; holding Control turns speech into text without the audio leaving the Mac. Command+H opens the full history, which can be searched by meaning rather than by exact words.
The hardware bar is real: Apple Silicon from the M1 MacBook Air onward, macOS 14, and 4.9 GB of free disk space on an 8 GB Mac or 8.6 GB from 16 GB up. Installation is a drag-and-drop from a .dmg with no administrator password; the app is sandboxed, notarised and hardened by Apple, so it can see neither your screen nor your files, and it updates itself automatically.
Three privacy modes coexist. Offline keeps everything local. Fluid Pro encrypts the conversation and sends it to inference servers the publisher owns and runs in RAM-only mode inside its own EU facility, with optional AES256 second-layer encryption so that even the OVH gateway cannot read the payload; the encryption, decryption and inference server code is published on GitHub under GPL v3. The third mode connects Fluid to an external provider of your choice, such as OpenAI or Anthropic, and the publisher says plainly that this is not private at all.
Fluid Pro runs DeepSeek R1 Llama 3.3 70B at a claimed 50-plus tokens per second, reachable through an API covering a subset of the OpenAI 1.0 specification with the model name Fluid-Latest. An account is a single number that serves as both login and password; payment goes through Fungies.io and Stripe by card, by cash sent through the post, or in cryptocurrency after manual checks. For companies, FluidBox AI is a preconfigured private server installed on the customer's premises from EUR 19,999.
What it does
- Answer questions in natural language, fully offline, with the model running on your own Mac
- Take voice dictation while you hold the Control key, without the audio being sent anywhere
- Assist with coding, a use the publisher names explicitly for developers protecting their codebase
- Search the web on your behalf through Fluid Pro, from version 0.4.1 onward
- Find an earlier conversation by meaning rather than by exact keyword
- Act as an inference backend through an API compatible with the OpenAI specification
- Switch answering style between Logical, Balanced and Creative
When to use Fluid / When not to
A quick filter to help you decide if Fluid is the right fit.
When to use Fluid
- Mac users on Apple Silicon who want a ChatGPT-like assistant that keeps working with no internet connection
- Developers who refuse to send proprietary source code to a third-party cloud
- Professionals bound by confidentiality — lawyers, clinicians, journalists — who need drafting help on sensitive material
- Privacy-minded individuals who would rather pay anonymously, in cash or cryptocurrency, than register an identity
- Small EU companies looking for a private AI server installed on their own premises
When not to use Fluid
- Windows and Linux users, and owners of Intel Macs: Fluid requires Apple Silicon and macOS 14
- Anyone needing a mobile or browser-based assistant — there is no iOS, Android or web client
- Organisations handling information of the highest sensitivity; the publisher itself points them to an air-gapped server instead
- Teams needing volume from the API, which is capped at 3,000 inference requests per month
- Buyers who require a signed Data Processing Agreement or a certification such as SOC 2 or ISO 27001 — none is published
How to use Fluid
A typical end-to-end flow, from setup to results.
- Check the requirements first: an Apple Silicon Mac, macOS 14 or later, and 4.9 GB or 8.6 GB of free disk space depending on your RAM
- Download the Fluid disk image from the publisher's CDN — downloading it counts as accepting the Terms and the Privacy Policy
- Open the .dmg, drag the app into the Applications folder and launch it from there
- Let the app download its Llama 3 brain on first launch; that download is what the disk space is for
- Press Ctrl+Space from any application to bring Fluid up, then type your question
- For voice, grant the recording permission at startup, enable Voice Recognition from the status bar icon, then press Ctrl+Space and hold Control while speaking
- Press Escape to stop a generation, or simply ask again to interrupt it
- Press Command+H to open the history and Command+N to start a new thread
- For Fluid Pro, create an account on the account site, save the account number — it is never shown again — then add time to it
- Test the API with a POST to the chat completions endpoint, an Authorization Bearer header carrying your account number, and the model name Fluid-Latest
Pros & Cons
Pros
- Privacy is structural rather than contractual: in offline mode nothing leaves the machine at all
- The Mac app is free in beta and needs no account to be created
- Zero-logging and zero-storage on the API, with inference servers deliberately run in RAM-only mode
- The encryption, decryption and inference server code is open source and auditable under GPL v3
- Anonymous payment is genuinely possible: cash sent by post, or cryptocurrency after verification
- One-click install with no administrator password, in an app sandboxed and notarised by Apple
- EU publisher under Polish law, with a Data Protection Officer named and reachable by email
Cons
- Apple Silicon macOS only: no Windows, no Linux, no Intel Mac and no mobile client
- macOS 14 minimum, plus 4.9 to 8.6 GB of free disk space for the local model
- Still a 0.x beta, absent from the Mac App Store, and outdated versions eventually stop working
- The client software is not open source; only the server components are published
- The API is capped at 3,000 inference requests per month, changeable without prior notice
- No Data Processing Agreement, no security certification and no published retention period
- Liability is capped at twelve months of fees, and access can be suspended at the publisher's sole discretion
Pricing & Plans
A permanent free plan is available: the beta version of the Mac app is offered at no charge. The lowest paid entry point is Fluid Pro, at USD 19 per month excluding VAT, cancellable at any time. A perpetual licence for the Mac app is also offered, covering the current version and six months of upgrades, but its price is not published. The FluidBox AI private server for businesses starts at EUR 19,999.
- Fluid App (beta) — free
- with inference running locally on the Mac
- Fluid Pro — USD 19 per month plus VAT
- 3
- 000 inference requests monthly
- and the premium plan of the Mac app unlocked at no extra cost
- Fluid App perpetual licence — one-time purchase covering the current version and six months of upgrades
- with all premium features
- price not published
- FluidBox AI — on-premises private server from EUR 19
- 999
- with tailored support plans
- offered to EU corporate customers
Data, GDPR & hosting
A consolidated view of how Fluid handles your data.
GDPR overview
GDPR is addressed explicitly. Section 6 of the Fluid Pro API privacy policy states that, as a Polish company, the publisher is committed to complying with the General Data Protection Regulation, and that a Data Protection Officer has been appointed and can be reached at legal@getfluid.app. Concrete rights are named: access and correction by email, deletion of the account at any time with all associated data permanently removed, and an opt-out obtained by cancelling the subscription or deleting the account. The three legal documents all carry an effective date of 17 January 2025, and Polish law governs the Terms. Two gaps are worth noting: no Data Processing Agreement is published or offered anywhere on the site, and no Article 27 representative is named — none being required, since the company is itself established in the European Union.
Who owns the data?
Under section 3 of the Terms of Use you retain all rights to the content you create with Fluid, whether it was generated on your own Mac or online through Fluid Pro, subject only to the Meta Llama 3 and Qwen 2.5 model licences, which you must read and accept. Ownership of the software runs the other way: section 7 states that Fluid keeps every intellectual property right in the app and the API, and forbids selling, redistributing or reverse engineering them. On the data side the publisher declares that it collects, stores and sends no user-identifiable data; where a card payment is made, it does have access to the account email address supplied through Fungies.io.
Reuse rights
You may reuse whatever Fluid produces without asking permission: the Terms grant you all rights to generated content, in offline and online generation alike. Two limits apply. The first is licensing — your use remains subject to the Meta Llama 3 and Qwen 2.5 licences. The second is accuracy: section 2 warns that generative AI may return incorrect, misleading or offensive information, and places fact-checking squarely on you. Section 14 lists prohibited uses, including illegal activity, malware, spam and phishing, impersonation and anything that disrupts the service. A one-time purchase grants a perpetual licence to the current version, with six months of upgrades included.
Data retention & training
Hosting summary
Hosting depends on the mode you use. Offline there is no hosting at all: the model and the conversation stay on your Mac. With Fluid Pro, inference runs on servers the publisher owns and controls in a facility located in the EU, with no third-party access, and those servers operate in RAM-only mode so that no conversation content is persisted. The public-facing gateway is hosted at OVH, which also stores non-conversational data such as usage statistics, transactions and licence expiry; when second-layer encryption is enabled, OVH cannot decrypt the message content passing through it. Model downloads and app updates come from third-party CDNs, which may log IP addresses. No specific hosting country is ever named — only the EU region — and the location of the OVH servers is not stated. The publisher is established in Warsaw, Poland, and Polish law governs the Terms.
Where Fluid works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting Fluid.
- Generative output can be incorrect, misleading or offensive, and the Terms put fact-checking entirely on you
- In Fluid Pro mode the privacy guarantee rests on the publisher's own statements: the server code is auditable, the way it is operated is not
- Even offline, web search and remote images send the search phrase and the image URL off your machine — a documented exception that is easy to forget
- Connecting Fluid to an external provider such as OpenAI or Anthropic voids the privacy promise entirely
- Crash reports and basic usage events go to Sentry, and switching that off only became possible in version 0.3.3; outdated versions eventually expire
- Paying in cash by post carries a real risk of interception, loss or damage, and such payments are never refunded
- An assistant one keystroke away invites quiet delegation of thinking, and a small self-funded publisher that names no director also carries continuity risk
Setup & Integrations
Technical difficulty
Very low for the Mac app: download a disk image, drag it into Applications and launch it. No administrator password is required, and the publisher advertises a one-click install needing no special skills. The only real hurdles are hardware — an Apple Silicon Mac and 4.9 to 8.6 GB of free disk space. Voice recognition adds three permission steps. Fluid Pro is a developer-level task: create an account, keep the account number, fund it, then issue an OpenAI-compatible HTTP call. FluidBox is installed on site by the publisher's team, training included.
Deployment
Integrations
Supported languages
Behind Fluid
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Fluid.
Frequently asked questions
Do my conversations leave my Mac?
What hardware do I need to run Fluid?
Is Fluid free?
Is my data used to train the models?
Is Fluid available on the Mac App Store?
Is Fluid open source?
Where is the data hosted when I use the API?
Can I pay without revealing my identity?
What is the usage limit on the API?
Is there a minimum age?
Should you pick Fluid?
Fluid answers a question most AI assistants dodge: where does the computation actually happen? On an Apple Silicon Mac the answer is your own machine, and in offline mode the conversation genuinely does not leave it. That single decision shapes everything else — the Ctrl+Space shortcut over any app, the local dictation, the modest but real bill of 4.9 to 8.6 GB of disk space, and the absence of any account to create.
Where the local model is not enough, Fluid Pro offers a paid path that stays unusually consistent with the pitch: RAM-only inference servers in a facility the publisher owns in the EU, optional AES256 second-layer encryption so the gateway host cannot read the payload, server code published under GPL v3, and cash accepted by post if you would rather not be identified at all. At USD 19 per month excluding VAT it undercuts buying and tuning a GPU machine, which is precisely the comparison the publisher invites.
The trade-offs are just as clear. The product is macOS-only on Apple Silicon, still numbered 0.x, absent from the Mac App Store, and capped at 3,000 API requests a month. The client itself is closed source, no Data Processing Agreement or security certification is published, and liability is capped at twelve months of fees. The publisher calls itself a small, self-funded business, names no director publicly, and openly tells anyone handling information of great importance to buy an air-gapped server instead.
Read as what it claims to be — a privacy-first everyday assistant for Mac users, with an EU-hosted API for heavier work — Fluid is coherent and unusually candid about its own limits. Read as an enterprise platform, it is not there, and does not pretend to be.
- Choosing a selection results in a full page refresh.
- Opens in a new window.