KoltrixOrganize your inbox by chat: labels, archive and stars safely
The six reversible actions an assistant has in Koltrix, a find-show-confirm-act-check pattern for bulk cleanups, how to undo mistakes and when rules do better.
On this page(8 sections)
Tidying an inbox is satisfying and tedious in equal measure. You know exactly what to do, label these, archive those, and it still takes forty minutes of clicking.
An assistant connected to Koltrix can do the clicking while you do the deciding. This post covers the six organizing actions it has, a safe pattern for bulk cleanups, how to undo anything, and when a standing rule is a better answer than asking an assistant every time.
The six things an assistant can change
With the Organise permission, a connected assistant can do these and nothing else to your mail:
| Action | Tool | What happens |
|---|---|---|
| Add a label | apply_label |
Adds the label to a conversation |
| Remove a label | remove_label |
Takes the label off; the label itself is kept |
| Archive | archive_thread |
Moves the conversation out of the inbox into Archive. Nothing is deleted |
| Mark read | mark_read |
Marks every message in the conversation read |
| Mark unread | mark_unread |
Marks it unread again |
| Star or unstar | star_thread |
Stars it, or unstars it |
Two features of that table matter. First, every action is reversible, and the tool descriptions say so: archive can be moved back from Archive in Koltrix, mark read reverses with mark unread, and a label comes off with remove_label. Second, the list is also the boundary. There is no tool that deletes a conversation and none that forwards one. Even a badly misled assistant can only shuffle your mail around.
You can see the same operations in the app, which is useful for checking. Nothing the assistant does is hidden: the labels, archive state and stars show up in Koltrix exactly as if you had clicked.
Labels: names, creation and typos
An assistant refers to a label by name or id. If the name you give does not exist yet, Koltrix creates it. That is convenient, since you can say "label these Receipts" without making the label first. It also means a typo makes a new label rather than raising an error.
Two habits prevent a mess:
- Ask for your existing labels first: "List my labels." The assistant has a tool for this (
list_labels), and it shows each label's name and color. - Use consistent names in your prompts, and prefer the exact name over a paraphrase.
If you do end up with a stray label, removing it from the conversations is a one-line request.
The safe pattern for a bulk cleanup
For anything affecting more than a handful of conversations, use five beats: find, show, confirm, act, check.
- Find. "Search for receipts from the last month: anything with 'receipt', 'invoice' or 'payment confirmation' in the subject."
- Show. "List what you found, sender and subject, before changing anything."
- Confirm. You scan the list. Remove anything that does not belong: "Skip the one from Lena."
- Act. "Label the rest 'Receipts' and archive them."
- Check. "How many did you label? Show me the first five."
The step that people skip is the show-and-confirm. It is the cheapest safeguard there is. Because organizing is reversible, a mistake here costs you a minute, not a customer. But a mistake that you notice in the preview costs nothing at all.
A worked example
Find newsletters from the last two weeks that I have not opened. List sender and subject. Do not change anything.
After the list:
Mark those as read, label them "Newsletters", and archive them. Skip anything from my accountant.
Three tool calls per conversation (read state, label, archive), made by the assistant, and a handful of seconds of your attention. You can stop at any step.
Searching precisely
Cleanup prompts depend on finding the right set. Koltrix search understands operators such as from:, to:, subject:, label:, has:attachment, is:unread, in:sent, before: and after:, and the search tool for assistants can also filter by folder, label, sender (a full address or a domain), and unread. You do not need to type them. But when a prompt returns the wrong set, naming the sender's domain or a date range usually fixes it:
Search for mail from acme.com before the 1st of this month.
Be specific about folders as well. The folders an assistant can look in are inbox, sent, drafts, archive, spam, trash and starred. If you meant only the inbox, name the folder in your prompt rather than leaving it to chance.
Things worth knowing
- One conversation at a time. Each organizing call acts on a single conversation. A cleanup of 60 conversations means many calls. It works, but each connection has its own request limit, so a very large sweep may be slower or need to be done in batches. Ask for it in chunks ("do the first twenty").
- Mark read is a real change. Reading a thread through the assistant does not mark it read, but the
mark_readaction does, for every message in it. If you rely on unread as a to-do list, do not ask the assistant to "clean up unread" without saying what counts. - Archive is not delete. Anything archived is in Archive and searchable. You can ask the assistant to bring it back by searching the archive folder.
Standing rules versus asking each time
An assistant is best for one-off sweeps. For mail that always gets the same treatment, a rule is better.
| Situation | Better tool |
|---|---|
| Every receipt from one sender should be labeled Receipts | An auto-label rule in Koltrix |
| Clean up last quarter's newsletters once | The assistant, using the pattern above |
| Decide which of this morning's threads matter | The assistant, because it needs judgment |
Koltrix has auto-label rules for the routine cases, and they run without anyone present. The assistant covers the cases a rule cannot express, the ones where you would otherwise have to read the thread to decide. Using both is common: rules for the predictable, chat for the rest. MCP, the API or an automation tool goes into how to choose.
Undoing a mistake
If a cleanup went further than you wanted, the undo is straightforward because every action reverses:
| Mistake | Undo |
|---|---|
| Wrong label on some conversations | Ask the assistant to remove it, or remove it in Koltrix |
| Archived the wrong ones | Move them back from Archive in Koltrix, or ask the assistant to find them in the archive folder |
| Marked read by accident | Ask it to mark them unread |
| Starred the wrong ones | Unstar them |
If you are not sure what it changed, auditing what an AI assistant did in your mailbox explains where those actions are recorded. Tool calls are logged in the workspace audit log with the tool's name, never the content of your mail.
Key takeaways
- An assistant with the Organise permission can label, unlabel, archive, mark read or unread, and star. It cannot delete or forward.
- Every action can be undone, so organizing is a comfortable first write action.
- A label name that does not exist yet gets created, so check spelling or list labels first.
- For bulk cleanups, use find, show, confirm, act, check, and do large sweeps in chunks.
- Use rules for mail that always gets the same treatment and the assistant for the rest. Overview of the integration.
Start with Koltrix
Your domain, one inbox, and an API that sends.
A team inbox where AI sorts and drafts (nothing is sent without your click), plus the transactional API and SMTP relay your product sends with. 7 days free, no card.
