Should I vibe code
AI email assistant that organizes inboxes, drafts replies, blocks cold email, and automates cleanup
An autonomous agent with delete permission on your inbox is a very fast way to lose something.
?
Their verdict, the Starter price and the build-time estimate come from their entry, MIT-licensed. Checked 2026-08-03.
?
Our verdict, the regret score and everything below it. Editorial and unsponsored — nobody can pay to be moved.
The honest answer
why the verdict is what it is
Reading and summarising your mail is fine. Letting a model archive, delete and reply on your behalf hands irreversible actions to a system that is confidently wrong a few percent of the time.
What actually breaks
not "if". the specific failures.
- An archive rule that matches far more than you meant it to
- A reply sent on your behalf, to the wrong person, saying something you did not say
- A model confidently classifying something urgent as noise
- OAuth tokens with full mailbox scope, stored somewhere you have not threat-modelled
The rule was “archive anything that looks like a newsletter”. It ran overnight across the full mailbox rather than the last 24 hours, and the classifier’s idea of a newsletter turned out to include your accountant. You do not know what is missing, because the only record of what was archived is the archive, and the thing you needed was time-sensitive.
Is that you?
the verdict is a default, not a law
- It only reads and suggests, and every action requires a click
- It operates on a label, not the whole mailbox
- Nothing it does is irreversible
- It can send, delete or archive without a human confirming each action
- It runs unattended on a schedule
- It has full mailbox scope rather than a narrow one
If you build it anyway
the checklist, then the prompt that enforces it
- Suggest, never act. A human approves every state change until you have months of evidence, and probably after that too.
- Request the narrowest OAuth scope that works. Never full mailbox access if read-and-label will do.
- Nothing deletes. Ever. Move to a label the model owns, and let a human empty it.
- Dry-run mode is the default, and it prints exactly what it would have done.
- Cap actions per run. A rule that wants to touch 400 messages should stop and ask.
- Log every action reversibly, with enough detail to undo it in bulk.
I want an AI assistant over my email. Build it under these rules: 1. Read-only by default. The first working version must be incapable of sending, deleting or archiving. Prove the classification is good before any write capability exists. 2. Request the narrowest OAuth scope that satisfies the feature. If I ask for full mailbox access, ask me what specifically needs it. 3. When write actions are eventually added: never delete. Move messages to a dedicated label that the assistant owns and that I empty manually. 4. Every batch action is capped (default 25 messages) and anything above the cap requires explicit confirmation showing me the full list first. 5. Dry-run is the default mode and prints a table of intended actions. Live mode requires an explicit flag on every invocation, not a config setting. 6. Append every action to a local append-only log with message IDs and the previous state, and write an 'undo last run' command in the same commit as the action it undoes. 7. Never auto-send. Drafts only. A reply the model wrote is a draft in my drafts folder that I open and send myself. 8. Store tokens in the OS keychain, not a dotfile. Show me the refresh flow. 9. In the README, list every irreversible operation the tool can perform. If that list is not empty, explain why each one earned its place.
That one keeps you out of trouble. For the prompt that actually builds it, canivibecodeit.com has one.
their build prompt ↗Or don’t build it
the boring option, and the way back out
If what you want is triage suggestions, the safe homemade version is genuinely good and worth building. If what you want is autonomous inbox management, the honest answer is that nobody has made that trustworthy yet, including the funded companies.
$20/mo is cheaper than your weekend.
Because the tool only ever labels and drafts, walking away costs nothing: delete the label, revoke the OAuth token. That is the whole point of building it this way.
The official open-source product, with a supported self-hosting path.
Questions
Why is this rated harsher than other AI tools here?
Because the actions are irreversible and the input is adversarial. Email is the one inbox where strangers can put text in front of your model, which makes prompt injection a delivery mechanism rather than a theoretical concern.
- GDPR Art. 5 — principles relating to processing of personal data
- GDPR Art. 32 — security of processing
Every week, someone ships something they shouldn’t have.
New verdicts, the worst thing that landed in the trap, and the occasional incident report. No other email, ever.
last reviewed 2026-08-03 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice