Proton releases Lumo Desktop: an AI agent with file and shell access, under your control
Proton, best known for its private email and cloud storage, has released Lumo Desktop for Windows, macOS and Linux. The AI assistant works with files, the terminal and services such as Jira. For each connection, you choose whether it acts on its own, asks permission first, or doesn't run at all. Chats are encrypted, and user data is not used to train models. For teams wary of giving an agent access to their disk, it is a solid template for permission settings.
AI-processed from TechCrunch; edited by Hamidun News
Desktop AI agents that read your folders and run commands in the terminal have become commonplace. So has the anxiety that comes with them: to do its job, such an assistant needs broad access to your disk, which means a risk of leaking personal data. The same week, Apple announced tighter settings for apps that request Full Disk Access — exactly the kind of apps in question. Against this backdrop, Proton has released a desktop version of its AI app Lumo, betting on control and privacy.
What Lumo Desktop can do
The app launched on Tuesday for Windows, macOS and Linux. Lumo used to be a chatbot. Now it can carry out tasks on the user's computer: work with files and folders, run commands in the shell, pull in content from the browser and automate routine work.
The set of connectors is fairly practical. The file system, shell and browser content retrieval are available locally. Among the services are Proton Mail, Confluence, Dropbox and Jira. The connectors support local and remote MCP servers, so you can plug Lumo into your own data sources and tools, just as you can with other agents.
There are also Skills, as in other AI apps. You save instructions, templates and reference files in them so you don't have to explain a task from scratch in every chat. Existing Custom Lumos will be migrated to the Skills format automatically.
How access permissions work
Proton's main difference from its competitors is the granularity of permissions. For each connector, the user decides what Lumo does on its own, what it does only after confirmation, and what is blocked entirely. A simple example: you can let the agent read a folder of reports but forbid it from deleting files, and allow any terminal command to run only with your explicit "yes." Individual tools can be switched off right in a conversation via the Tools menu.
The same logic applies to scheduled tasks. Say you ask Lumo to compile a summary from Jira and Confluence every morning. That task gets its own set of permissions: what it does automatically, and what it stops and waits for your confirmation on. This way a background job doesn't get more authority than you gave it.
Privacy and infrastructure
Proton promises that conversations and data from connected apps and files are not used for personalized advertising or for training models. Chat history is encrypted so that only the user can read it, not the company. Proton keeps no server logs, and the limited data that does remain is protected by Swiss privacy law. Hosting is in Europe, outside the reach of the US CLOUD Act, which governs how law enforcement obtains data from cloud providers.
At the same time, the company is expanding its compute capacity: the first NVIDIA B300 cluster in Switzerland, built to Open Compute Project standards, is being launched. The source gives no details on which models run on this hardware or what share of Lumo's workload it will take on.
What this means in practice
Developers and product managers can take away two points. First, a permission model of "automatic, with confirmation, forbidden" at the level of each connector and each scheduled task is gradually becoming the expected standard for agents with disk access. If you are building your own agent, compare your settings with this scheme. Second, MCP server support makes Lumo suitable for integration into work processes, not just for personal use.
There are caveats. The claims about encryption, the absence of logs and data being kept out of training remain the company's promises: the article cites no independent verification. Pricing is not stated, and neither is how Lumo compares with ChatGPT and Claude on the tasks themselves. An agent with shell access remains a risky tool under any permission model, so it makes sense to start with folders that contain nothing critical and to keep confirmation switched on for dangerous actions.
- Nov 20, 2026Amazon SageMaker Now Flags Idle GPUs and Rising Latency on LLM Endpoints on Its Own
- Nov 20, 2026OpenAI Agent Wrote a Script to Secretly Copy Email Instead of a Health Document
- Nov 20, 2026OpenAI opens Daybreak to NATO and the EU: 95% of 88 reported bugs turned out to be real
- Nov 20, 2026AWS teaches AI agents to buy APIs in Marketplace and pay through a regular AWS bill
Key Questions Answered
What Lumo Desktop can do?
The app launched on Tuesday for Windows, macOS and Linux. Lumo used to be a chatbot. Now it can carry out tasks on the user's computer: work with files and folders, run commands in the shell, pull in content from the browser and automate routine work. The set of connectors is fairly practical. The file system, shell and browser content retrieval are available locally. Among the services are Proton Mail, Confluence, Dropbox and Jira.
How access permissions work?
Proton's main difference from its competitors is the granularity of permissions. For each connector, the user decides what Lumo does on its own, what it does only after confirmation, and what is blocked entirely. A simple example: you can let the agent read a folder of reports but forbid it from deleting files, and allow any terminal command to run only with your explicit "yes." Individual tools can be switched off right in a conversation via the Tools menu. The same logic applies to scheduled tasks.
What are the key details on Privacy and infrastructure?
Proton promises that conversations and data from connected apps and files are not used for personalized advertising or for training models. Chat history is encrypted so that only the user can read it, not the company. Proton keeps no server logs, and the limited data that does remain is protected by Swiss privacy law. Hosting is in Europe, outside the reach of the US CLOUD Act, which governs how law enforcement obtains data from cloud providers.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.