Proton Lumo Desktop shipped November 17, 2026 for Windows, macOS, and Linux — a privacy-first desktop agent with Connectors, Schedules, Skills, and MCP support, backed by a public GPLv3 repo and Swiss B300 inference infrastructure.
The launch lands beside Google’s free Gemini Tasks and OpenAI’s GPT-6.2, but Proton’s pitch is different: give agents access to files and shell, while keeping chats encrypted, never training on connector data, and letting users gate every tool.
Primary sources: Proton’s Lumo Desktop blog, lumo-desktop on GitHub, and Proton’s download page.
What Proton announced
Eamonn Maguire’s November 17 post introduces three builder-facing surfaces:
- Connectors — filesystem, shell, browser fetch, Proton Mail Bridge, Google Drive, Jira, Confluence, and arbitrary local or remote MCP servers.
- Schedules — recurring or one-off tasks with per-task permissions (auto, ask, block).
- Skills — saved instructions migrating from older Custom Lumos formats.
Proton also confirmed NVIDIA B300 capacity on OCP hardware in Switzerland, positioning Lumo as sovereign European AI rather than U.S.-jurisdiction inference.
TechCrunch’s same-day piece highlighted Full Disk Access tensions on macOS — Apple has been tightening security prompts when AI apps request broad disk reads, a thread we noted near Apple Siri limit betas.
Lumo Desktop architecture (from the open repo)
Tri-state approval gate before Lumo connectors can run shell or file tools
The README describes a Tauri shell around the Lumo web client:
| Feature | Behavior |
|---|---|
| Approvals | Each tool: auto, ask, or disabled |
| Quick ask | Global shortcut opens a search window |
| Scheduled prompts | Cron-style runs with stored permissions |
| Credentials | OAuth tokens and MCP secrets in OS keychain |
| Sign-in | Browser-based auth so passkeys work |
| Updates | Signed auto-update channels per platform |
Packages ship as universal macOS DMG, Windows MSI/exe, and deb/rpm for recent Ubuntu and Fedora.
For MCP-heavy teams, this is a reference desktop host you can diff against Hugging Face Agent Manager patterns and Blender MCP add-ons.
Connectors: what Lumo can touch
Proton’s connector catalog mirrors enterprise agent design:
- Local compute: read/write files, shell execution, browser content retrieval.
- Email: Proton Mail via Bridge — aligned with Proton’s engineering manifesto on internal AI use.
- SaaS: Jira and Confluence for ticket and doc search.
- MCP: user-supplied servers plus presets (GitHub, Linear, Figma, Grafana, Stripe, PubMed per README).
Each connector gets tri-state permissions. Example pattern Proton gives: allow inbox search, ask before send, block deletes. That is the same governance shape as Bedrock human-in-the-loop designs, but on a consumer/practitioner desktop.
Schedules and Skills: recurring agent work
MCP connectors bridging Lumo Desktop to files, mail, and third-party tools
Schedules accept natural language (“weekly inbox summary”) or manual cron-like setup. Pausing, editing, or deleting tasks does not require wiping global permissions — each schedule carries its own policy object.
Skills store reusable instructions and reference files — Proton’s answer to Gemini Skills replacing Gems and OpenAI Custom GPT patterns, but with exportable templates under user control.
Privacy claims builders should verify
Proton repeats commitments relevant to security reviewers:
- No training on chats or connector content.
- No ads personalized from Lumo data.
- Zero-access encryption on history — Proton cannot read chats.
- Swiss law and non-U.S. CLOUD Act hosting story for stored metadata.
Those are policy statements, not a penetration test. Teams under GDPR or banking rules should still run DLP scans on shell connectors and block unvetted MCP servers, same as we advise after GTIG backdoor research.
Lumo Desktop vs ChatGPT / Claude desktop clients
| Question | Lumo Desktop | Typical Big Tech AI desktop |
|---|---|---|
| Source available? | Yes (GPLv3) | No |
| Per-tool approvals? | Granular tri-state | Varies; often coarse |
| Email inside privacy envelope? | Proton Mail Bridge | Vendor mail cloud |
| Inference geography | Switzerland (B300) | Vendor default regions |
| Consumer task mode | Schedules + connectors | e.g. Gemini Tasks |
Lumo is not trying to win FrontierBench; it tries to win trust per action.
How to try Lumo Desktop safely
- Download from proton.me/lumo/download.
- Start with filesystem read-only connectors before enabling shell.
- Register MCP servers in a VM if you do not trust remote tool authors.
- Review scheduled tasks weekly — stale prompts with auto-approve are risky.
- Mirror the same approval discipline on any Codex or Claude Code agents you run beside Lumo so one desktop client does not become the only gated tool surface.
Step-by-step: first hour with Lumo Desktop
- Install from Proton’s download page and sign in with browser OAuth so hardware keys work.
- Open Connectors and enable filesystem read only; run a test question on a harmless folder.
- Add Proton Mail Bridge if you use Proton mail — without Bridge, email connectors stay idle.
- Register one MCP server you control (not a random GitHub preset) and set every tool to ask.
- Create a Schedule for a weekly inbox summary with send actions still set to ask.
- Migrate one Custom Lumo into Skills and confirm the template appears in a fresh chat.
That sequence mirrors how we onboard Hugging Face Agent Manager users: narrow blast radius first, widen tools after review.
IT and compliance talking points
Procurement teams will ask how Lumo differs from ChatGPT desktop with Full Disk Access. Proton’s answers are encryption, no training, European hosting, and inspectable source. You still need local policy for shell execution — Lumo can run commands your DLP stack must monitor.
Swiss B300 investment is as much geopolitical positioning as latency optimization. If your org forbids U.S. inference for certain matters, Lumo is a documented alternative; if you only care about benchmark scores, it is not.
What the launch does not solve
- No telephony equivalent to Gemini’s dealership calling in today’s Proton post.
- Connector catalog smaller than Google’s 60+ consumer integrations.
- Model frontier unspecified in the Proton post — privacy, not leaderboard scores, is the sell.
Compared with OpenAI and Google desktop agents this week
On November 17, 2026, three personal-agent stories collided: GPT-6.2 for paid work intelligence, Gemini Tasks for free consumer chores, and Lumo Desktop for privacy-first connector control. Picking one is less important than documenting which data leaves which trust zone.
| Product | Trust pitch | Best for |
|---|---|---|
| Lumo Desktop | Swiss encryption, GPLv3 client | Developers who want MCP with per-tool ask gates |
| Gemini Tasks | Google VM isolation + AP2 | Families already on Gmail, Maps, Wallet |
| ChatGPT Work + 6.2 | OpenAI frontier quality | Teams already standardized on Codex |
Lumo does not win raw Agents’ Last Exam scores. It wins when legal or healthcare policy forbids training on connector content and when IT wants to read the Tauri connector code before rollout.
Skills migration from Custom Lumos
Proton says older Custom Lumos migrate into Skills. Export any legacy Lumo configs before upgrading desktop builds so you do not lose templates mid-migration. Skills behave like Gemini Skills replacing Gems: reusable instructions, not one-off system prompts buried in chat history.
FAQ-adjacent questions IT will ask
Can Lumo run without internet? The desktop shell can reach local files and shell offline, but model inference still needs connectivity to Proton’s servers unless Proton ships an on-device tier later. Plan network policy accordingly.
Does Lumo replace Claude Code? No. Lumo is a connector-aware assistant. Claude Code remains a coding agent with its own sandbox story — see Claude HTML in Canvas for Anthropic’s parallel desktop push. Many engineers will run both: Lumo for mail and files, Codex or Claude for git.
Why GPLv3? Proton explicitly wants community inspection of connector behavior. Security reviewers can diff releases on GitHub the same way they audit FFmpeg skill repos trending on Hacker News.
Hardware requirements? Tauri builds target current Ubuntu/Fedora and universal macOS; older Linux distros may need the AppImage or manual build from source listed in the README.
Lumo vs Proton Mail only? Proton still sells Mail, Drive, and VPN separately; Lumo Desktop is an additive agent surface. Your Proton subscription tier may gate usage limits described on the Lumo download FAQ — confirm quotas before scheduling heavy MCP workloads. Free accounts can start desktop trials, but connector-heavy schedules may need paid Lumo limits sooner than chat-only use. Watch Proton’s changelog for connector preset updates — MCP catalogs change quickly.
Inspecting the GPLv3 repo before you trust connectors
Because lumo-desktop is public, security teams can review how Tauri exposes shell and filesystem APIs before any employee installs the MSI. Clone the repo, trace the approval gate implementation, and compare it with your internal software allowlist. The README’s preset MCP list (GitHub, Stripe, PubMed, etc.) is a menu of high-risk tools — block presets you do not recognize, then expand deliberately.
Proton’s browser-based sign-in is a deliberate choice so passkeys work; document that flow for IT help desks so users do not paste passwords into the embedded webview.
