DEV Community

Cover image for Proton Lumo Desktop Is Here: A Privacy-First AI Agent That Can Actually Do Things
Paul Lovelace
Paul Lovelace

Posted on Originally published at toolscout-e2i.pages.dev

Proton Lumo Desktop Is Here: A Privacy-First AI Agent That Can Actually Do Things

Proton has spent the last year telling anyone who will listen that its AI assistant, Lumo, is different: no training on your chats, no ads built from your data, everything under Swiss privacy law. On November 17, 2026, it took the next step and gave Lumo a proper home on your computer. Lumo Desktop is now available for Windows, macOS and Linux, and it is not just a chat window in a nicer frame. It comes with three features that turn Lumo from something you talk to into something that does work for you: Connectors, Schedules and Skills.

That is the interesting part. The desktop app itself is expected. The agent capabilities are the story.

The short version

Lumo Desktop lets the assistant reach into your files, run shell commands, fetch web content, connect to Proton Mail, Google Drive, Jira, Confluence and Dropbox, and talk to local or remote MCP servers. It can run tasks on a schedule, each with its own permissions. And it can remember reusable instructions, templates and reference files, applying them automatically when they are relevant. Proton says it is publishing the source code for these tools so people can inspect how they work, and you need a Proton account to use the app.

If you have been waiting for a desktop AI agent that does not treat your data as the price of admission, this is the most serious attempt yet. It is not perfect, and the privacy claims deserve the same honest scrutiny as any vendor's. But the design choices here are genuinely different from the defaults set by ChatGPT and Claude.

Connectors: Lumo can finally touch your work

Connectors are the headline feature, and the one developers should look at first.

Until now, Lumo was a chatbot you pasted things into. Connectors change the direction of travel: Lumo can now reach out and pull information from the places you already work. The local tools include filesystem access, shell execution and browser content fetching. The connected services include Proton Mail (through Proton Mail Bridge), Google Drive, Jira, Confluence and Dropbox, plus local and remote servers that speak the Model Context Protocol (MCP).

That last item matters more than it looks. MCP is the open standard Anthropic introduced for connecting AI assistants to tools and data sources, and it is fast becoming the common plug for this kind of integration. Supporting it means Lumo Desktop can, in principle, connect to any MCP server you already run, local or remote, rather than only the integrations Proton ships. For developers who have been building MCP servers for Claude, that is a low-friction second home for the same tooling.

The part Proton got right, and the part every other desktop agent should copy, is the permission model. For each connector action you choose one of three settings: Auto-approve, Ask human, or Deny. This is per action, not per connector, and scheduled tasks carry their own permissions on top of that. It is a small detail that changes the whole risk profile. An agent that can run shell commands on your machine is powerful; an agent that can run shell commands only when you said so, for that specific action, is usable.

The honest caveat: shell execution plus filesystem access is the most dangerous combination any desktop AI can have. The permission model only protects you if you actually configure it instead of clicking Auto-approve on everything during setup. Treat the first hour with Connectors like onboarding a new junior engineer with production access: give it the smallest scope that lets it do the job, and widen it later if you must.

Schedules: tasks that run while you are not there

Schedules let Lumo run a task at a set time or on a recurring basis. You can describe the job in plain language or enter the instructions and timing manually, and each scheduled task has its own permissions, so a weekly inbox summary does not inherit the same approval rules as a monthly spending review.

The obvious use cases are the boring ones, which is exactly the point: a weekly digest of your inbox, a Monday-morning summary of open Jira tickets, a monthly check of a folder of receipts. The value of a scheduled agent is not that it does something spectacular; it is that the boring thing happens without you remembering to ask.

Two things to watch. First, a scheduled task that can act on its own is only as safe as its permissions, and permissions set once and forgotten are how incidents happen. Proton lets you pause or delete scheduled tasks, so build the habit of reviewing them the way you review cron jobs. Second, Proton has hinted that Connectors and Schedules could come to Lumo mobile in the future; for now, this is a desktop-only capability.

Skills: your instructions, reusable

Skills are reusable instructions, templates and reference files that Lumo applies automatically when they are relevant, and it can use several in one conversation. If you have used custom instructions in other assistants, the idea will feel familiar. The difference is in the packaging: a Skill can bundle the instruction with the reference material it needs, so "write release notes in our format" can carry the format template with it instead of relying on you to paste it every time.

Existing Custom Lumos, Proton's earlier version of saved system prompts and document references, are being migrated to Skills automatically. If you already built a library of Custom Lumos, you should not have to rebuild it.

For teams, this is the quiet productivity feature. A shared Skill that encodes how your team writes a bug report, triages a support thread or drafts a changelog is worth more than any single clever prompt, because it compounds every time someone uses it.

The privacy claims, stated honestly

Proton's pitch is that connected information and conversations are never used to train models or serve personalised ads, that chat history is protected with zero-access encryption, and that the service runs on European infrastructure under Swiss privacy law, outside the scope of the US CLOUD Act. It is also publishing the source code for these tools for inspection, and the launch was announced by Eamonn Maguire, Proton's Director of Engineering for AI, as the first release under the company's pledge to share AI tools it built for its own staff.

Here is the honest version: those are Proton's own statements, not the findings of an independent audit of this release. That does not make them false. Proton has a long public record on privacy, Lumo itself is open source, and publishing the tools' source code for inspection is a real accountability step that most competitors do not take. But a privacy claim is only as strong as its verification, and the right way to read this launch is as a strong design plus a promise, not a proven guarantee. If your threat model includes nation-state access or regulated data, read the source and the privacy policy yourself before connecting anything sensitive.

What is genuinely different, and verifiable in the architecture, is the local-first shape of it. Connectors that run against your local filesystem and local MCP servers keep more of the work on your machine than a cloud-only agent does. That is a real structural advantage, not just a policy paragraph.

How it compares to ChatGPT and Claude on the desktop

ChatGPT and Claude both have desktop apps, and both have moved toward agentic features. Claude has computer use and MCP support; ChatGPT has scheduled tasks and deep integrations of its own. On raw capability, Lumo Desktop is not ahead of them today. It is narrower: fewer integrations, a smaller model ecosystem, no equivalent of Claude's computer-use breadth yet.

Where Lumo is ahead is on the axis Proton has always competed on. Neither OpenAI nor Anthropic offers zero-access encryption on chat history as a default, neither excludes your data from training with the same absoluteness, and neither is built under Swiss privacy law with open-source code you can inspect. Lumo 2.0, launched a few months ago, added Lite and Max models, memory across conversations, image generation and a side-by-side editing canvas, so the underlying assistant is no longer the minimal chatbot it started as in July 2025. The gap is closing on features while the privacy gap stays open.

The practical question is the one every developer asks about a new tool: is it good enough for the work I actually do? For writing, summarising, triaging mail and tickets, and running scheduled digests, the answer is increasingly yes. For the deepest coding-agent workflows, Claude and ChatGPT still have the edge. Lumo Desktop is the better choice when the data you would feed the agent is the thing you care most about protecting.

What developers should do with it

  • Try the permission model first. Install it, connect one low-risk source such as a folder of public documents, and set everything to Ask human. Live with that for a week before auto-approving anything.
  • Write one Skill for a task you repeat. A release-notes template, a bug-triage checklist, a weekly status format. That is where the compounding value lives.
  • Stand up a local MCP server and point Lumo at it. If you already have MCP tooling for Claude, this is nearly free, and it tells you quickly how serious Proton is about the open standard.
  • Schedule one boring task. A weekly inbox digest is the canonical first schedule. If it works, add the Jira summary. If it does not, you have learned the failure mode cheaply.
  • Read the source Proton is publishing. The whole point of open-sourcing the tools is that someone checks them. Be that someone, at least for the connectors you enable.

Pricing and availability

Lumo Desktop is available now for Windows, macOS and Linux from Proton's download page, and it requires a Proton account. There is a free tier. Lumo Plus costs $9.99 per month (billed at $119.88 annually) and adds unlimited questions, unlimited Projects and unlimited Skills, with a 30-day money-back guarantee. Proton is also bringing a new NVIDIA B300 computing cluster online in Switzerland to strengthen its AI infrastructure, which suggests this is the start of a bigger push rather than a one-off release.

The bottom line

Lumo Desktop is the first desktop AI agent that treats your permission as a first-class feature rather than a settings-page afterthought, and the first from a company whose entire brand is not training on your data. The capabilities are real but young, the privacy story is strong but partly unverified, and the shell-access permission model will punish anyone who configures it carelessly. For developers who have been waiting for an agent they can point at their real work without handing their data to an ad business, this is the launch to try first.

Verdict: Install it, lock the permissions down, and give it one real recurring job. If it earns your trust there, it has earned a place on your desktop.


Sources: Proton's Lumo Desktop announcement, November 17, 2026, as reported by Neowin, Cyber Insider, Thurrott, TechBriefly and MacSources; Proton AI engineering director Eamonn Maguire's announcement remarks; Lumo 2.0 background from Neowin's coverage. Privacy claims in this article are Proton's own statements, not an independent audit. Pricing: Lumo Plus $9.99/month, billed annually at $119.88, per Proton's pricing page as reported November 18, 2026. No hands-on testing was performed.

Top comments (0)