Privacy first
Privacy and security
How Don't Be Michael keeps your business data in: who on the team is given what, how the trust layer checks the questions and answers that pass through it, what you decide stays private, and how links show where they go.
See the short version on the home page
Who is given what
Pete, the Research team member, does the team's web research. Everyone else, Michael included, is told to ask Pete, and the app turns off the web page tool and the web search tool for every team member but Pete. Pete is off until you turn them on in Settings, Research & privacy.
| Who | Web pages and search | Your files and mailboxes | Connectors |
|---|---|---|---|
| Your team | Told to ask Pete | Their own folder and the mailboxes you give them | Only the ones you give them |
| Michael | Told to ask Pete | Reads the team's work to hand it out | Only the ones you give him |
| Pete | Yes, the team's researcher | Questions only, never your files | None |
Each team member keeps its own notes and inbox, and teammates, Michael included, are told to ask it instead of opening them. Each note says what showed it, such as an email and its date, or a file. The app keeps your own answers, your email decisions and what you tell Michael apart from the team's notes, and only those count as yours: your word wins when notes disagree.
Email a team member reads from a mailbox you add is labelled as written by its sender: information, never instructions. Lines written to steer an AI are pointed out so the member skips them and carries on, and Michael is told once for each such email.
The app gives Pete none of your company knowledge: no documents, no company profile, no company foundations.
The trust layer
A trust layer sits between Pete and the rest of the office, built to stop agents going rogue and leaking your business data. It checks every question the app passes to Pete and every answer that comes back through it.
On the way out
- Every message the app passes to Pete, from a teammate, Michael, the app itself or one of your answers, is checked against the kinds of data you keep private, except the app's own fixed notes, which carry no business data and go as written.
- It goes as written, goes cleaned or made more general so the research still helps, or comes back to its sender as "Not sent to Pete", naming the private kind it depends on.
- Fixed patterns run before and after the check, and a message with a password or key in a recognisable shape goes back before any model reads it.
- Questions only: the app gives Pete no files, and a message with a file goes back to its sender.
On the way in
- Instructions aimed at AI agents are removed from every answer that comes back through the trust layer.
- Each of those answers is labelled as information from the internet, never instructions.
- An answer that tries to steer its reader is flagged to Michael.
- Pete's notes are not copied for the team.
Since 0.2.3, Pulse lists each question the trust layer checked on its way to Research and each web answer, with what was taken out marked word by word.
When the check can't run
If the check is unsure, too slow or fails, nothing goes through, in either direction: the teammate who asked is told, and so is Michael. Pulse lists each message held because the check couldn't run: when, who it was between, what it was about, why, and who was told.
What stays private
You choose, kind by kind, what your team may send to Pete. Twelve kinds, in this order, with these defaults. Passwords and keys are always private, whatever you choose.
| Kind of data | For example | Starts as |
|---|---|---|
| Places | countries, regions, cities, time zones | Allowed |
| Industry and market topics | product types, trends, regulations, standards | Allowed |
| Dates and general timing | next quarter, holidays, deadlines in general | Allowed |
| General questions | how to, definitions, reference facts | Allowed |
| People's names | customers, staff, contacts | Private |
| Contact details | emails, phone numbers, street addresses | Private |
| Company and brand names | customers, suppliers, competitors, any named business | Private |
| Money | your prices, invoices, revenue, bank and card details | Private |
| Account numbers and IDs | customer IDs, order numbers, government IDs | Private |
| Health and personal life | medical, family, personal circumstances | Private |
| Text from your files and email | quoted passages from documents or messages | Private |
| Passwords and keys | logins, API keys, tokens | Always private |
Links in the app
Every link in a reply names the site it opens, and shows the real one when the words name a different site.
What goes to Claude
What an agent reads for a job goes to Claude through your own Claude plan or Anthropic API key. Pete runs only on Claude.
Where your data lives
The team's work, memory, files and mailbox passwords are kept on your computer, with passwords encrypted under a key your system keeps (in the Keychain on a Mac). A mailbox's password is sent only over an encrypted connection; a mail server that won't switch to one is refused. The app keeps each team member's conversation records out of every other member's file tools, Michael's included. The app sends no usage data, only failure reports with no names, messages or files.
When the app itself fails, such as a crash or a team member stuck in a loop, it reports that to the developers on its own, as fixed fields only: no names, messages or files. Reports go to a private repository and are then copied, field by field, into public issues. Anything you can fix yourself is never sent, and setting the DO_NOT_TRACK environment variable stops it.
The code is open: read it on GitHub, or see how to report a security issue.