Lumo Desktop: Set Boundaries Before Connecting Apps

Laptop displaying app permission settings for a Lumo Desktop connector workflow

You want an AI assistant to find a file, summarize an email, and prepare a follow-up without giving it free rein over your laptop and business accounts. Proton’s Lumo Desktop now offers file-and-app automation on Windows, macOS, and Linux. The useful question is not whether it can connect to your tools; it is what each connection is allowed to do.

What Lumo Desktop adds—and who should try it

Lumo Desktop is worth evaluating if your work moves between local files and connected services and you want to control the assistant’s access. The desktop launch introduced three building blocks: Connectors for access, Schedules for timed tasks, and Skills for reusable instructions. Those features make it more than a chat window, but they also make setup decisions consequential.

Proton announced Lumo Desktop on November 17, 2026 for Windows, macOS, and Linux. Its Connectors can access local files, execute shell commands, and fetch browser content. Lumo can connect to Proton Mail through Proton Mail Bridge; Proton also lists third-party connectors including Slack, Atlassian, Dropbox, and Google Drive. Custom local and remote Model Context Protocol (MCP) servers are supported as well.

For a solo developer, that could mean asking Lumo to inspect project notes and draft a status update. For a small business operator, it could mean gathering context from a document and a message thread before preparing a reply. Neither example requires the assistant to send the reply or edit the source material. Start by separating finding information, drafting a proposed action, and taking the action. They carry different risks.

The desktop app is a reasonable fit when you can name a narrow task and the sources it needs. It is a poor starting point if your plan is simply to connect every account and see what happens. Access design is part of building the workflow, not cleanup after it runs.

What Proton’s privacy claims do—and do not—cover

Proton says saved Lumo chat history uses zero-access encryption and that it does not use conversations or connected-app and file data for model training or personalized ads. That is relevant if you handle business correspondence. It does not mean prompts stay on your computer: Proton says prompts are sent to its servers over TLS for processing.

That distinction matters because “private” can describe several different things:

Question What the cited Proton guidance establishes What you still need to decide
Are saved chats protected? Proton says saved chat history uses zero-access encryption. Whether a particular document should be submitted for processing at all.
Does the model train on my work? Proton says it does not use conversations or connected-app and file data for training or personalized ads. Whether your own client or company rules permit use of the service.
Does processing happen only on my device? No. Proton says prompts go to its servers for processing using TLS. Which content you are comfortable sending for that purpose.
Can the assistant reach connected tools? Yes, when connectors are enabled and their actions are permitted. Which tools and actions the task actually needs.

Privacy protections and access controls solve different problems. A service can have strong protections for stored chats while an over-permissioned connector still reads the wrong file or performs an unwanted action. Treat Lumo’s privacy statements as inputs to your decision, not as a substitute for reviewing the task, data, and permissions.

For sensitive work, make a small test set first: a non-sensitive file, an ordinary message, and a clearly defined output you can inspect. If the task only needs a summary, do not use real client records to discover whether the assistant can summarize.

Configure connectors as permissions, not conveniences

The most important Lumo setup fact is easy to miss: Proton’s connector documentation says actions default to Auto-approve, and connected connectors are enabled in every conversation by default. You can change each connector action to Auto-approve, Ask human, or Deny, and turn connectors off for individual conversations. Do not assume a new connection is approval-first.

Before connecting an app, write down the task in one sentence: “Read the shared project notes and draft an update; do not post or change anything.” Then configure permissions against that sentence. If an action is unrelated, deny it. If it changes an external system, consider human approval. Reserve automatic approval for actions you have a reason to allow without review.

A practical starting policy looks like this:

Action type Starting choice Reason
Read a known source needed for the task Ask human while testing Confirm that the assistant selects the intended source.
Search or read broad business data Ask human or Deny A useful answer may not justify broad access.
Send, post, edit, or delete Ask human or Deny These actions can affect customers or records outside the chat.
Execute shell commands Deny unless essential Commands can have effects beyond reading a file.
Access an unrelated service Deny No task requirement means no access requirement.

This is a suggested review policy, not a description of Lumo’s defaults or a guarantee that every connector exposes identical actions. Inspect the actions actually presented for the connector you use.

Pay particular attention to shell access and custom MCP servers. A shell command is not equivalent to reading a document; its effect depends on the command and the environment in which it runs. A custom server can make additional tools available to the assistant. Before enabling either, identify who maintains it, what resources it can reach, and whether the task works without it. If you cannot answer those questions, leave it disconnected.

Finally, check the connector state in each new conversation. Because connected connectors are enabled by default across conversations, “I only used this for one task” is not the same as “it is unavailable for the next task.” Turn off connections a conversation does not need.

Test one file-and-app workflow before automating it

A safe first test has a fixed input, a visible proposed output, and no automatic external change. For example: use an ordinary project document and a related message to draft a follow-up, then have a person verify the names, commitments, and recipients before anything is sent. This tests whether Lumo finds the right context without making a customer-facing mistake.

Build the test in this order:

  1. Define the outcome. Specify the artifact you want: a draft reply, a short summary, or a list of unresolved items. “Handle this account” is too vague to evaluate.
  2. Select only the required sources. Connect the file or service needed for the test. Leave unrelated connectors off in that conversation.
  3. Review action permissions. Check every relevant read and write action rather than relying on the default Auto-approve setting.
  4. Supply a boundary in the task itself. Tell Lumo which material to use, what not to infer, and that the result is a draft for review.
  5. Inspect the result against the sources. Look for omitted context, invented commitments, and content pulled from an unintended place.
  6. Test an exception. Use a document with missing information and see whether the output flags the gap instead of guessing.

You could give the assistant an instruction like this:

Draft a reply using the project notes and the related message. List any missing facts separately. Do not send a message, edit a file, or treat an inferred deadline as confirmed.

That instruction makes the desired behavior clear, but it is not an access control. Keep connector permissions aligned with it. If the task says “do not send” while a send action remains on Auto-approve, you have left a preventable conflict in the setup.

Record what happened in plain language: which sources were used, which permissions were needed, what the assistant got wrong, and whether the draft was usable after review. You do not need a benchmark to make a decision about your own workflow. You need a repeatable test and a clear threshold for trusting the next step.

If the task fails, narrow it before adding more access. A request to “summarize all customer activity” may fail because “customer,” “activity,” and the source of truth are undefined. Adding another connector will not fix those definitions.

Use Schedules and Skills only after the manual task works

Lumo’s Schedules can run a task at a specified time or repeatedly, with task-specific connector permissions. Skills hold reusable instructions and can include templates or reference documents. Together, they can make a tested task repeatable; they cannot make an ambiguous task reliable.

First get a manual run to produce a useful, reviewable result. Then use a Skill to preserve the instructions that made it work: the output format, which source takes precedence when records disagree, and how to mark missing information. A Skill is a good home for a drafting template. It should not be treated as proof that connected data is accurate or that an action is authorized.

Only then consider a Schedule. For a recurring project update, the intended flow might be:

Read the permitted project sources
→ produce a draft update
→ flag missing or conflicting facts
→ leave external posting to a person

This is a workflow design, not a claim about a built-in Lumo template. It deliberately ends before an external action. When you set up the scheduled task, review its task-specific connector permissions rather than assuming the permissions from a manual conversation will express the same boundary.

Scheduled work also needs an owner. Someone must know what output to expect, where to review it, and what to do when a source is unavailable or a draft is wrong. If nobody will check the result, scheduling a task can turn an occasional mistake into a recurring one. Keep the scheduled scope smaller than the manual test until its outputs are consistently useful for your team.

Decide on Free, Plus, or another approach

Lumo Free is a practical place to test a narrow workflow, but do not choose a paid plan until you know which limits your real task reaches. Proton lists Lumo Free at $0, and its getting-started guidance says Free includes one Project. Proton advertises Lumo AI Plus at $9.99 per month when billed $119.88 every 12 months and says Plus has higher usage limits than Free.

Proton’s terms also say “unlimited” usage on plans including Lumo Plus remains subject to “fair and reasonable use.” Do not interpret a paid plan as unrestricted automated or programmatic capacity. Check the current pricing page and terms against the way you intend to use it before committing.

Your need Sensible next step
Evaluate drafting from a small set of sources Start on Free and test permissions and output quality.
Use Lumo often enough to reach Free’s limits Review Plus limits, billing terms, and fair-use conditions.
Run unattended, business-critical actions Test the full failure and review process before relying on a desktop assistant.
Require strict limits on what data an assistant can reach Define those boundaries first; compare tools against them rather than buying on a privacy label alone.

Proton lists Lumo 2.0 Lite, Lumo 2.0 Max, and Apertus 1.5 as model options. That gives you options to test on your own documents, but the cited material does not establish which will perform best for your workload. Use the same representative task and check the output yourself.

The decision is less “Can Lumo replace ChatGPT or Claude?” than “Can I give this assistant the specific context it needs while keeping the rest out?” For file-and-app work, that is the test that matters.

An access-boundary checklist

The approach proposed in this article is to set explicit access boundaries for a file-and-app task: what it needs to read, what it may propose, and what a person should approve before an external action. The same checklist applies whether the interface is a desktop assistant or another automation tool.

If you are evaluating Lumo for a business workflow, begin with one recurring task and its required sources. Before connecting your accounts, define the task and the access it needs.


Work with BizFlowAI

If you'd rather have this built for you, that's what we do: production AI automation for solo founders and small teams — agents, integrations, and document pipelines that actually ship.

Request a free discovery call — 30 minutes, we map the highest-ROI automation in your workflow. No pitch deck, just engineering.

More guides like this on the BizFlowAI blog.

Frequently asked questions

Are Lumo Desktop connectors approval-first by default?

No. Proton’s connector documentation says actions default to Auto-approve, and connected connectors are enabled in every conversation by default. Review each action and set it to Auto-approve, Ask human, or Deny before using a connector. Turn off connectors that a particular conversation does not need.

Does Lumo Desktop process my prompts entirely on my computer?

No. Proton says prompts are sent to its servers over TLS for processing. It also says saved chat history uses zero-access encryption and that conversations and connected-app data are not used for model training or personalized ads. Those protections do not remove the need to decide whether a document is appropriate to submit.

How do I safely test a Lumo Desktop file-and-app workflow?

Start with a non-sensitive file, a related message, and a specific output such as a draft reply. Enable only the sources the task needs, review connector permissions, and require human approval for actions that could change or send anything. Compare the draft with its sources and check whether Lumo flags missing facts instead of guessing.

Should I let Lumo Desktop run shell commands or use a custom MCP server?

Only if the task requires that access and you understand what it can reach. Shell commands can affect more than the file you want to read, while a custom MCP server can expose additional tools. Identify who maintains the server, what resources it can access, and which actions need approval. Leave either capability disconnected if you cannot answer those questions.

When should I use Lumo Desktop Schedules and Skills?

First run the task manually with narrow sources and permissions, then review whether its output is reliable. A Schedule runs a task at a specified time, while a Skill provides reusable instructions. Neither replaces connector permissions or review of actions that could send, edit, or delete information.