Industry·
Proton Brings Lumo to the Desktop With Granular Agent Permissions
Proton's Lumo now runs as a desktop agent on Windows, macOS and Linux, with per-connector permissions, MCP support and Swiss-hosted inference — a template for privacy-first agentic AI.

Proton has moved Lumo, its privacy-positioned AI assistant, off the browser tab and onto the desktop, releasing builds for Windows, macOS and Linux. The shift matters less for the interface than for the capability: Lumo can now act on local files, execute shell commands, fetch browser content and reach into connected services, turning a chat product into a task-performing agent.
Permissions are set per connector, not per app
The desktop client ships with connectors for the local file system, shell execution and browser fetching, alongside Proton Mail and third-party services such as Confluence, Dropbox and Jira. It also speaks to local and remote MCP (Model Context Protocol) servers, the emerging interface agents use to reach external data. For each connector, users choose whether actions run automatically, require explicit approval, or are blocked outright; individual tools can be switched off mid-conversation, and scheduled tasks carry their own separate permission sets. Chat history is encrypted so that only the user can read it, no server-side logs are retained, and stored data sits on European infrastructure under Swiss privacy law, outside the reach of the U.S. CLOUD Act. A Skills feature, replacing the earlier Custom Lumos, lets users save reusable instructions, templates and reference files.
Why it matters for GPU / AI infrastructure: agentic desktop clients multiply tool-call round trips, and every one of those is a latency-sensitive inference request rather than a single chat completion. MCP is quickly becoming the integration contract that capacity planning has to account for, because connector-heavy workloads produce bursty, long-context traffic that is poorly served by batch-oriented provisioning. Just as telling is Proton's claim to run its own compute: the company says its first NVIDIA B300 cluster, built to Open Compute Project standards, is now live in Switzerland. Data-residency promises are only as credible as the silicon behind them, and for buyers in the UAE and wider Gulf region the lesson is direct — sovereignty arguments in AI procurement now arrive with hardware attached, and inference locality is becoming a line item rather than a marketing phrase.
- aigpu
- ai gpu
- ai gpu cloud
- aigpu dubai
- ai agents
- mcp
- data privacy
- inference infrastructure
By AiGpu Editorial · Editorial rewrite based on public reporting (TechCrunch AI)
← All articles