
As AI assistants evolve from chatbots into agents that read files, manage email and act on a user’s behalf, they need deeper access to sensitive data and, increasingly, the authority to make decisions. Proton’s latest release is an attempt to show that this shift does not have to come at the expense of privacy or control.
The company has launched Lumo Desktop for Windows, macOS and Linux, introducing three capabilities to its privacy-first assistant: Connectors, Schedules and Skills. Together, Proton says, they turn one-off AI interactions into repeatable workflows.
An agent stack built on familiar patterns
The architecture will be recognisable to anyone following the agentic AI space. Connectors give Lumo access to data and tools: the local filesystem, shell execution, browser content fetching, Proton Mail via Proton Mail Bridge, Dropbox, Jira and Confluence. Crucially, Lumo also supports custom local or remote Model Context Protocol (MCP) servers, plugging it into the fast-growing ecosystem around that standard.
Schedules handle autonomy over time, letting Lumo run tasks at set times or on a recurring basis. Users can describe a task in natural language and Lumo will ask for any missing details before preparing it for review. Skills provide persistent context – saved instructions, templates and reference files that Lumo applies automatically when relevant, with several able to run in the same conversation. Users can write their own, ask Lumo to help build one, or upload an existing Skill. Existing Custom Lumos will migrate to the new format automatically.
In Proton’s framing, Connectors supply the information, Schedules decide when work happens, and Skills define how it should be done.
Human-in-the-loop by design
Where Proton is trying to differentiate is governance. Each connector and each scheduled task has its own permission set, allowing users to mark individual actions as automatic, approval-required or blocked. Lumo might be free to search an inbox, for example, but must ask before sending an email and cannot delete messages.
“As AI assistants become more capable, people should be able to benefit from that progress without giving up control over their personal data,” said Eamonn Maguire, Director of Engineering for AI at Proton.
“With Lumo Desktop, we’re bringing AI into everyday workflows while keeping privacy and user choice at the centre,” he added. “Connectors, Schedules, and Skills let people work with their apps and files, automate tasks, and personalize how Lumo works for them — while deciding what it can access and which actions require their approval.”
Transparency as a feature
Proton is also publishing the source code for Lumo Desktop, which it calls the first step in giving its community more visibility into how the company uses AI, including tools it builds for internal use. The new features began life as internal tools to let Proton’s own teams use AI within strict privacy and security requirements.
On data, Proton maintains that conversations and connected-app data are never used to train models or serve personalised ads. Chat history is protected by zero-access encryption, Lumo keeps no server-side logs, and the limited data it does store sits on European infrastructure under Swiss law, which Proton says puts it beyond the reach of the US CLOUD Act.
For an industry still working out how much autonomy to hand AI agents, Proton’s approach – granular permissions, inspectable code and a strict no-training policy – offers one model for how capable assistants and user control might coexist. Lumo Desktop is available now.



