What Is Entropic? Quai’s Private AI Assistant, Tested (2026)

Most AI assistants ask you to trust them. Quai Network’s Entropic let us verify. We installed it, watched every connection it made for an hour, and put it through real work. For a privacy-first desktop assistant, the things that matter held up.

Why Test a Private AI Assistant the Hard Way?

Every AI assistant on the market makes some version of a privacy promise. Almost none of them let you check it. So, in order to test Entropic, we did what any sceptical engineer would do: we put a network monitor on it and watched exactly what left the machine, connection by connection, for an hour of real use.

That is worth saying up front because, as far as we can find, nobody has tested Entropic before. This is the first, and we ran it the way we would test a tool we were about to trust with our own work.

In short, the privacy architecture is real, and the design decisions behind it hold up to scrutiny.

What Stays on Your Machine?

This is the part that matters, and it is where Entropic earns its “local-first” description.

Your files stay put. We placed a document on the machine and asked Entropic about it. The agent runs inside its own sandboxed workspace that is walled off from the host system, so it does not roam your disk, index your folders, or read anything you have not deliberately handed it. In an era where plenty of “assistants” ask for full-disk access on day one, an agent that starts with no view of your files at all is a deliberate and welcome design choice. You decide what it sees.

It runs silently when idle. We left the app open and untouched for five minutes with the monitor running. Zero new outbound connections. No background chatter, no telemetry drip, nothing phoning home while you are not using it. That is rarer than it should be.

Code runs locally. We asked it to write and execute a Python script. It did, correctly, inside its isolated container, and the execution generated no new external connections. The sandbox is not a marketing word here; it is doing real work, keeping code off your actual system while it runs.

Voice and transcription are on-device. The speech and audio models are marked “Private” and run locally, so dictation and read-aloud do not depend on a cloud service at all.

Add to that two things a sceptic will want to know. There is no crypto involved: no wallet, no token, no blockchain account, despite Quai’s roots. And the code is open source under the MIT licence, published on GitHub, which is more transparency than closed-core rivals like LM Studio offer. When we did sign in, the Google consent screen asked only for name, profile picture and email. Nothing more.

What Leaves Your Machine, and Why?

Here is the honest and simple part, and it is true of every capable AI assistant, Entropic included.

When you actually ask the model a question, that prompt travels to the model provider. We watched it happen: a request went out to infrastructure hosted on Amazon Web Services, the model did its work, and the answer came back. That is not a leak, it is how any tool built on GPT-class models functions. The intelligence lives in a data centre; talking to it means sending it your question.

What Entropic gets right is the boundary. Your files, your workspace and your code stay on your machine. Only the message you choose to send, when you choose to send it, goes out, and it goes to the provider you select rather than to some opaque middle layer. Entropic’s product page describes this arrangement plainly.

And if even that round trip is more than you want, Entropic gives you a lever most competitors do not: a “bring your own keys” mode. Plug in your own OpenAI, Google or OpenRouter credentials and your prompts go directly to a provider you already have a relationship with, on your own account, with Entropic stepping out of the middle. For a regulated business that already has an enterprise agreement with a model provider, that is exactly the control you would want.

The practical takeaway for a privacy-conscious user is that far more stays on your device with Entropic than with a browser tab pointed at a chatbot, and the one thing that does leave is the thing that has to. Where that line sits between local and cloud is the whole question sovereign-AI buyers are now asking.

Does Entropic Actually Work?

We ran five real tasks a working professional would throw at it. It handled them well.

For the “find a current legal date with a source” test, we asked for the entry-into-force date of a regulation that took effect two days before the test. Entropic got it right, which means it is reaching live sources rather than leaning on stale training data. For anyone doing current-awareness work, that is the difference between a useful tool and a confident wrong answer. When its search step hit a configuration hiccup, it told us so and went to primary legal sources instead of guessing, which is exactly the behaviour you want from a tool you are trusting with facts.

The file and document work was the standout. Hand it a note or a pair of PDFs and it returns a clean, accurate summary in under a minute, every specific detail intact. This is the everyday use case for most people, and it is where Entropic feels most finished.

Where Does Entropic Fall Short?

Entropic is early. It reports itself as a 0.1.x build, and it behaves like capable early software: the free credits are generous but finite, and switching into bring-your-own-keys mode needs a valid key configured before it will start, which is logical once you know it but can catch you out. The bundled runtime is a sizeable first download. None of this is unusual for a young product moving fast, and none of it touched the core experience once set up.

It’s important to note that Quai markets Entropic as v1.0, while the build we tested reported itself as v0.1.21. We call it early software on the strength of that build number and how it behaves, not the label on the download page.

The file workflow is upload-based rather than the agent reaching into your folders directly. Given the security payoff, walling the agent off from your disk, we think that is the right trade for most users, but it is worth knowing going in: you hand files to it, it does not go looking.

Who Is Entropic For?

Entropic fits the person or team that wants a capable AI agent without handing over the keys to their whole machine. The sandboxed workspace, the local code execution, the silent idle behaviour, the open-source core and the bring-your-own-keys option add up to a tool built by people who took the privacy question seriously rather than pasting it onto the box.

For a business weighing AI tools against where data goes, Entropic is a credible option worth piloting, especially in bring-your-own-keys mode, and especially because you can verify its behaviour yourself rather than taking a vendor’s word for it. That verifiability is, in the end, the most valuable thing it offers.

The Verdict: A Private AI Assistant That Survives Scrutiny

We went in to check whether Entropic’s local-first pitch survived contact with a network monitor. It did. Your files stay yours, your code runs on your own hardware, the app does not chatter when idle, and the one thing that leaves, your prompt to the model, is under your control and can be routed through your own provider account. It is early software with the usual early edges, but the foundations are the right ones, and it is rare to find a product in this space that invites this kind of scrutiny and comes through it well.

The best compliment we can pay Entropic is the plainest: we tried to catch it out, and it earned our trust instead.

This article is for information only and is not financial advice. Tested on macOS on 29 July 2026 using Entropic v0.1.21, monitored with Little Snitch Network Monitor and lsof.

Author: Akos Szima

See also:

What Happened to Le Chat? Mistral’s Vibe Rebrand, Tested (2026)

What Is Bielik? Poland’s Sovereign AI Model, Tested (2026)