Proton's Lumo Assistant Moves to the Desktop, With Per-Connector Permission Controls

Proton has launched Lumo Desktop for Windows, macOS, and Linux, extending its privacy-focused AI assistant from the browser onto the desktop, where it can carry out tasks involving local files, connected applications, and automations. TechCrunch reported the launch on 17 November 2026.
The desktop client ships with three built-in connector types: file-system access, shell execution (running commands directly in the operating system's terminal), and browser content fetching. Lumo also connects to Proton's email application and to third-party services including Confluence, Dropbox, and Jira. Connectors support both local and remote Model Context Protocol (MCP) servers. MCP is an open standard that lets AI agents reach external data sources without custom integration work on each side.
Permissions are set per connector. Users can configure each one to allow actions automatically, require approval for each action, or block the connector entirely. Individual tools can be disabled from the Tools menu, and scheduled tasks can be created with their own permission settings. Skills let users save instructions, templates, and reference files for reuse across conversations.
On data handling, Proton states that conversations and information drawn from connected apps and files are never used to serve personalized ads or to train AI models. The company says it encrypts Lumo chat history so that only the user can read it, not Proton itself. It also says it keeps no server-side logs for Lumo and stores limited data under Swiss privacy law on European infrastructure.
Proton is also building its own compute. Its first NVIDIA B300 cluster, a group of data-centre GPU servers built on Open Compute Project standards (openly published hardware designs), is going live in Switzerland. The cluster is described as in progress.
In my view, the most consequential detail is the permission model rather than the connector list. A shell-execution connector set to allow actions automatically carries a very different risk profile from the same connector set to require approval. The outcome of a misconfigured agent, or one manipulated by prompt injection (hidden instructions planted in content the agent reads), depends heavily on which default a user accepts. The per-action approval option and the ability to disable individual tools will largely determine how the feature is used in practice, and they put that decision with the user rather than the vendor.
The broader context here is that AI assistants have been moving from chat windows toward agents that act on a user's machine and accounts. Lumo's pitch rests on the claim that this capability and a strong privacy posture can coexist, backed by data handling commitments that Proton has stated itself. Those commitments are Proton's to make, and they can be checked against the product's behaviour over time. For users who already keep sensitive correspondence, documents, and project tracking inside Proton and the tools Lumo connects to, a local agent operating under explicit, per-connector rules is a practical option that did not exist in this form before. The open question for technical buyers is how the approval workflow and the MCP connector model hold up under sustained real-world use, and that will become clearer over the coming months.


