Proton launched Lumo Desktop for Windows, macOS, and Linux on November 17, 2026, adding Connectors, Schedules, and Skills to its privacy-focused AI assistant. The release moves Lumo beyond answering questions: it can access connected information, execute permitted actions, and repeat work on a schedule, according to Proton’s announcement.
The immediate uses are practical. Lumo can gather reservations from email, review receipts across local folders and Google Drive, or prepare a weekly inbox summary using saved writing instructions.
The important distinction is between protecting information and authorizing actions. Proton says it will not train models on conversations or connected data. That promise matters, but it addresses a different risk from allowing an AI agent to execute shell commands or send email. Lumo’s controls deserve a close look before connecting sensitive accounts.
What Lumo Desktop Can Connect To
Connectors provide Lumo with access to information and tools. Proton’s launch announcement lists the following options across its three supported desktop platforms:
| Connector | Announced capability | Practical consideration |
|---|---|---|
| Local filesystem | Access files on your computer | Connected files may contain sensitive material unrelated to the task |
| Shell | Execute shell commands | Commands can have consequences beyond producing an answer |
| Browser content | Fetch browser content | Retrieved pages become material the assistant processes |
| Proton Mail | Connect through Proton Mail Bridge | Mail access depends on the Bridge connection |
| Google Drive | Work with personal files | Cloud documents add another source of potentially sensitive information |
| Jira and Confluence | Connect workplace tools | Company policies still apply to connected work data |
| Local or remote MCP servers | Connect tools using Model Context Protocol | Available capabilities depend on the server |
These are the capabilities described by Proton, not independently tested connector behavior. Browser content fetching, for example, should not be read as confirmation of unrestricted browser control.
MCP support makes the feature set extensible. A server can expose additional information or tools to the assistant, so the permissions of that particular server become part of the decision to connect it. “Supports MCP” is not a complete description of what an integration can do.
There is also a discrepancy in the launch coverage: TechCrunch lists Dropbox, while Proton’s announcement names Google Drive, Jira, and Confluence. I would use Proton’s list as the starting point and check the app before planning a Dropbox workflow.
Permissions Separate Searching From Sending
Lumo lets users assign actions within a connector to three categories: automatic, approval-required, or blocked. Proton illustrates the distinction with email:
| Setting | Proton’s email example | Consequence |
|---|---|---|
| Automatic | Search the inbox | Lumo can perform the action without asking each time |
| Approval required | Send an email | Lumo must request permission before sending |
| Blocked | Delete messages | The action is unavailable to the assistant |
This is a useful design choice because an email connector is not one indivisible permission. Searching messages, sending replies, and deleting mail carry different consequences.
For an inbox-summary workflow, I would allow searching, require approval for sending, and block deletion. The reasoning is straightforward: summarizing mail requires reading it, but it does not require changing or removing messages. A narrower configuration preserves the useful part of the workflow.
One default deserves particular attention. Proton states: “Connected tools are enabled in conversations by default.” Individual connectors can be disabled through a conversation’s Tools menu. That means users should review both the actions a connector permits and whether that connector belongs in the current chat. Proton’s announcement explains both controls.
The same logic becomes more consequential with shell access. Requiring approval helps only if the person approving understands what the proposed command will do. A connector’s usefulness is not a reason to automatically authorize every action it exposes.
Schedules Handle Timing; Skills Preserve Instructions
Schedules let Lumo perform a task once at a specified time or repeat it regularly. Users can describe the task and timing conversationally, then review and confirm what Lumo prepares. They can also create the task manually.
Each scheduled task has its own permissions. Its instructions and timing can be changed, and the task can be paused or deleted, according to Proton.
That separate permission layer is important. A weekly inbox summary and an automated email-sending workflow may use the same mail connection, but they should not necessarily receive the same authority.
Execution while the computer is asleep or the desktop app is closed remains unclear from the launch material. Before relying on a time-sensitive schedule, check those operating conditions. A task configured for a particular time is only useful if its execution requirements match how you use the computer.
Skills store reusable instructions, templates, and reference files. Lumo automatically applies relevant Skills and can use several within one conversation. Users can write a Skill, ask Lumo to help create one, or upload an existing one. Existing Custom Lumos will migrate automatically.
The three features fit together neatly:
- Connectors supply information and permitted actions.
- Schedules determine when work happens.
- Skills describe how the output should be prepared.
A monthly spending review shows the distinction. Connectors provide receipts, the schedule sets the review date, and a Skill specifies spending categories and the report format. Saving a Skill is a way to reuse context, not evidence that the model has learned a new capability.
Proton’s getting-started guide lists unlimited Skills under Lumo Plus. Its plan comparison does not provide a complete connector-and-schedule entitlement breakdown, so readers should check current account limits before building around those features.
Privacy Promises Cover Different Risks From Tool Access
Proton says Lumo does not use conversations or information from connected apps and files for model training or personalized advertising. It also says signed-in chat history uses zero-access encryption, meaning Proton cannot read that stored history. These are vendor claims, not independently audited findings in the sources reviewed here. Proton’s support documentation describes the history protection and training policy.
Those promises address important questions: how conversation history is protected and whether submitted information becomes training material. They do not make every authorized action safe.
Consider three distinct boundaries:
- Storage: Who can read saved chat history?
- Processing: What information can Lumo retrieve for a task?
- Action: What can Lumo change, send, or execute?
Zero-access encryption addresses the first boundary. Connector selection affects the second. Action permissions govern the third. Training policy is another separate choice, a distinction also relevant to .
Frequently Asked Questions
4 questions
1Which Platforms Support Lumo Desktop?
Proton launched Lumo Desktop for Windows, macOS, and Linux on November 17, 2026. The cited support page does not provide usable minimum OS version numbers.