Use cases
What people actually ask for.
17 things worth asking an assistant to do with a real mailbox, what each one does to the mail, and roughly what it costs in calls. Four of them are scheduling, and need a calendar connected. One of them is a thing no native connector on the market can do at all.
- Use cases 17
- Tools 28 email, up to 21 calendar
- Free tier 5 calls a day
Editorial
Written from a mailbox, not from a feature list
Every request on this page is one somebody would actually type, and every answer describes what happens to the mail rather than what the product supports.
The temptation with a page like this is to list what the tools do and put a question mark in front of each one. That produces one use case per tool, none of which anybody wants. What people actually want falls into four shapes: getting through a morning, answering something, finding something they half remember, and checking something a mail client cannot tell them.
Each card below carries a rough call cost, which is a number nobody else publishes. It matters because every plan here has a daily ceiling, so "about two calls" and "about ten calls" are the difference between a habit you can afford on the free tier and one you cannot. They are honest ranges rather than a specification: how many calls a request costs depends on how the model chooses to work, and two assistants answering the same question will not spend the same number.
Triage
Getting through the morning
The most common thing anybody does with this, and the one where the detail about read state earns its keep.
-
Go through my unread mail and tell me which of these actually need me, and which are just noise.
A morning triage pass that leaves nothing marked read
It lists the unread messages, opens the ones it needs to judge, and gives you a shortlist. The part that matters is what it does not do: reading never sets the seen flag, so every message it looked at is still unread when you open Outlook. Your own marker for "I have not seen this" survives the pass intact.
That is not an incidental nicety. An assistant that marks eighty messages read while triaging them has destroyed the only signal you had, and there is no undo at that scale.
Costs About 2 to 10 calls- list_emails
- read_email
- search_emails
-
Flag anything from a client that is waiting on me, and file the rest of this week's newsletters.
Flagging and filing in two calls, not two hundred
The write tools take a set rather than one message at a time: up to 500 messages per call for flagging, marking read, moving and deleting. So a filing session over three hundred newsletters is one call, not three hundred.
A move uses the IMAP MOVE command, so read state, flags and the original dates all survive it. The message arrives in the new folder looking exactly as it did in the old one, which is what makes the result indistinguishable from having done it yourself.
Costs About 3 to 6 calls- search_emails
- flag_email
- move_email
-
Anything in here I have already dealt with? Mark those read so I can see what is left.
Marking read as a deliberate act
Because reading does not mark anything, marking read is a separate tool that runs only when you ask for it. That is the right way round: the destructive-to-your-attention operation is the one that needs asking for, and the harmless one is free.
It is reversible too. mark_unread takes the same set of up to 500, so a pass you did not mean can be undone in one call.
Costs About 2 to 4 calls- search_emails
- mark_read
- mark_unread
Writing
Answering and sending
Where the difference between an assistant that drafts and one that actually works your mailbox shows up.
-
Attach the invoice from Hollis's email last Tuesday to a new message for the accountant.
A file moved between messages without passing through the model
This is the one nothing else on this market does the same way, and the mechanism is the reason. The assistant never sees the file. It reads the message, gets a handle to the attachment, and passes the handle to the send: the bytes go from your mailbox to the outgoing message on our side, and the model only ever holds a filename and a size.
So the only ceiling left is the message itself: 10 MB in total, the same figure any mail client applies. A 6 MB PDF costs no more in the conversation than a 6 KB one, which is the part that was impossible before. Compare that with Anthropic's Microsoft 365 connector, whose own security guide says attachments are not supported in any write tool, and that sending, forwarding and drafting all reject a message carrying one.
Costs About 2 to 3 calls- read_email
- send_email
-
Send the signed contract on my desktop over to the client.
A file off your own computer, without opening the panel
The file is not in the mailbox and has no web address, so there is nothing to hand the send. The assistant asks for a one-off upload link and then takes whichever of two routes it can: if it is able to run commands or make requests, it uploads the file itself and you do nothing at all. If it is not, it gives you the link, you open it and choose your files, and it carries on. Either way you never sign in anywhere.
What arrives goes into your own Drafts folder as a draft, which you can open, keep or delete in Outlook like any other. Nothing is stored on our side and the link stops working after half an hour. From there it behaves exactly like the case above: the assistant gets a handle rather than the bytes, so a 6 MB file costs no more in the conversation than a 6 KB one, and the only ceiling is the ordinary 10 MB on the whole message.
Costs About 3 to 4 calls- create_upload_link
- check_upload
- send_email
-
Draft a reply to this, but do not send it. I want to read it first.
A real draft, in the real Drafts folder
It is APPENDed to the mailbox's own Drafts folder, so it appears in Outlook, in webmail and on your phone, and you can edit it there, send it there, or delete it. A draft you can only see in a chat window is a record of a draft rather than a draft.
You can also ask for another go at it, which is the part most assistants get wrong. Say the second paragraph is too stiff and it rewrites that draft. It does not save a second one and it does not put the first in your bin, so a wording you worked through five times leaves exactly one message in Drafts. Anything already attached stays attached.
When you do send from here instead, exactly one copy is filed in the real Sent folder. Not none, which leaves you with no record, and not two, which is what happens when a server files a copy and the mail host files another.
Costs About 2 calls- read_email
- draft_email
- update_draft
-
Reply to this and tell them the meeting has moved to Thursday.
A reply that stays in its own thread
It replies inside the original conversation, quotes the original below the new text the way a mail client does, files the copy in Sent, and marks the original as answered, so your own client shows the little reply arrow next to it.
One safety rule worth knowing: a Reply-To header pointing somewhere the message did not come from is refused until the caller names that address explicitly. That header is how a redirected reply ends up somewhere you did not intend, and it is set by the sender rather than by you.
Costs About 2 calls- read_email
- reply_email
-
Forward the whole thread about the roof to the surveyor, with the photos.
A forward that carries its attachments
The original's attachments and inline images travel with it, under the forwarded-message block a real mail client produces. If a carried file will not fit inside the message ceiling, it is reported as skipped by name rather than being dropped quietly, so you find out at the point you can still do something about it.
Recipients are exactly the addresses supplied. Nothing is inferred, and nothing is added.
Costs About 2 to 4 calls- read_thread
- forward_email
Finding
Finding things
The category where a person is worse than a machine, because you remember the shape of a message rather than its subject line.
-
Find the thread with Hollis about the March invoice. It was some time last spring.
A search across folders, with the vagueness intact
One call searches up to 25 folders, by sender, recipient, subject, body, date range, read state, flag state and whether there is an attachment. The assistant turns "some time last spring" into a date range and a set of folders; you do not have to.
Results are capped, and this is the part worth trusting: the response reports the total found, the number returned, and whether it capped. So an answer built from the first fifty of four hundred results says so rather than reading like the whole story.
Costs About 1 to 3 calls- search_emails
- read_thread
-
What is the email address for the woman at the surveyors? I think her name is Ada something.
An address resolved from your own history
It looks the person up in the mailbox's own correspondence rather than in a contacts list you have not maintained since 2019. An address you have sent to outranks one that merely arrived from somebody using that name, because a display name on inbound mail can say anything at all.
That ranking rule is a small anti-phishing measure hiding in a convenience feature. The address you have actually written to is the one you trusted before.
Costs 1 call- find_contact
-
Read me the whole conversation about the tender, from the beginning.
A thread in one call, oldest first
From any single message in it. You do not have to find the first one, and the assistant does not have to make twelve calls to walk the chain.
A long thread is capped and the result says so, rather than quietly returning the newest few and letting a summary be built from the wrong half of a conversation.
Costs 1 call- read_thread
Checks
Checks a mail client cannot do
Two tools that are not mail-client features at all, and are the ones people are most surprised to find here.
-
Did that email to the council actually arrive, or did it bounce?
Bounces found and read, including the ones in Junk
It looks for delivery failure reports in the Inbox and in Junk, which is where a good proportion of them land, and it tells you whether each failure is permanent or temporary. That distinction is the whole point: resending fixes a temporary failure and simply duplicates a permanent one.
A bounce report is a machine-readable message that most people never open because it looks like noise. Having something read them for you is one of the plainest wins here.
Costs 1 to 2 calls- check_bounces
-
Is my domain set up properly so my mail gets trusted?
SPF, DKIM, DMARC and MX, checked and explained
It checks the connected mailbox's own domain and tells you what is there. It says when a DMARC record is set to p=none and is therefore publishing a policy that does nothing, which is the single most common way a domain looks configured and is not.
It also keeps a failed lookup apart from a missing record, because "we could not check" and "it is not there" are different problems with different fixes, and conflating them is how people spend an afternoon adding a record that already existed.
Costs 1 call- check_deliverability
Needs a calendar connected
Scheduling, on a calendar you already keep
These become possible once a calendar is connected, which is a separate step from connecting the mailbox. On Microsoft 365 that is approving calendar access on the same sign-in. On any other mailbox there are two routes and either will do: sign in once to connect Google Calendar, or give the address of your calendar server, which Fastmail, iCloud, Nextcloud and most hosts run. Two of the five below are Microsoft 365 only and say so where they appear.
-
Find an hour next week when Priya, Tom and I are all free, and book it with a Teams link.
A time everybody can make, found and booked in one ask
It reads everybody's free and busy blocks, picks an hour inside all three working days rather than one that is merely empty at seven in the evening, creates the meeting with an online link, and sends the invitations. The part worth watching is what it does when it cannot see somebody: that person comes back as unknown rather than free, and it says so instead of booking over an afternoon it could not read.
This one needs a Microsoft 365 mailbox with calendar access approved. It is also the case that most clearly beats doing it yourself, because the manual version is four calendars open in four tabs and a message that starts "does anyone mind if".
Costs About 3 to 5 calls- check_availability
- find_meeting_times
- schedule_meeting
-
What have I got on Thursday, and is there anything I have double-booked?
A day read back in your own timezone, repeats included
It reads the window you asked about and reports each event with who organises it and how many people are on it, so a block of your own time is never read back as a meeting. A weekly meeting appears on the date it actually falls on, the way it does in Outlook, rather than as a rule somebody has to interpret.
Every time comes back in the mailbox's own timezone and is stated with it. That sounds pedantic until the first time somebody answers "three o'clock" from a laptop in a different country. Works on any connected calendar, whoever hosts it.
Costs 1 to 2 calls- list_calendars
- list_events
-
I cannot make Tuesday. Decline it and suggest Wednesday afternoon instead.
A decline that carries an alternative, not just a no
It declines the invitation and proposes the new time in the same action, with a note to the organiser. That is the answer a person actually wants to give, and it is one most integrations cannot express at all: they can accept and they can decline, and the useful third option is missing. Where the organiser has switched proposals off, it says so rather than trying and failing.
Microsoft 365 mailboxes only. The organiser sees exactly what they would see if you had done it in Outlook, because it is the same request.
Costs 1 to 2 calls- read_event
- propose_new_time
-
I am off from Friday to the 14th. Set my out of office and move Monday's review.
Out of office set for a period, and the meeting moved
It writes the out-of-office reply for the exact date range, with a different message for people outside your organisation, then reschedules the meeting and tells everybody who was on it. The reply is the same setting Outlook writes, so it shows in Outlook and switches itself off at the end of the period.
Nothing of ours is running while you are away. Microsoft applies the setting at the time, exactly as it would if you had ticked the box yourself. Microsoft 365 mailboxes only.
Costs About 2 to 4 calls- set_out_of_office
- update_meeting
Accounting
What all this costs in calls
One MCP call is one tool run. The counts above are what the work on this page actually spends.
Reading costs calls exactly like writing does, which surprises people. A search is a call, opening a message is a call, and an assistant that reads three results before answering has spent four. That is why the free tier of 5 calls a day per mailbox is enough to try anything on this page and not enough to do it every morning.
The writes are cheaper than they look, and that is worth knowing before you budget for them. Flagging, marking read, marking unread, moving and deleting all take a set of up to 500 messages in a single call, so a filing session over three hundred newsletters costs one call rather than three hundred. A heavy tidy-up is one of the cheapest things you can do here.
The control panel shows your own real figures per mailbox, by day and by tool, over 90 days. That is the number to plan against rather than anything on this page, because it is yours. Pricing states both ceilings.
Every tool on this page is on the free tier. What you get less of is calls: 5 a day per mailbox, no card.
Limits
What people ask for that this cannot do
Asked often enough to be worth answering here rather than in a support reply.
-
Attach a file without being asked for it
A file already in the mailbox or at a web address attaches on its own, inside the ordinary 10 MB message ceiling. One on your own computer needs a one-off upload link first, which the assistant asks for and then either sends the file to itself or hands to you. A file the assistant composes has to be far smaller, because the model produces those bytes as output.
-
Send later, or on a schedule
There is no scheduled send. A message goes when the tool runs. If you want it to wait, draft it and send it yourself, which at least means a person read it before it left.
-
A calendar you have not connected
Connecting a mailbox does not connect a diary. Approve calendar access on Microsoft 365, or on any other mailbox sign in to Google or give the address of your calendar server, and the calendar tools appear; until then you get the email tools and are handed no calendar tools rather than tools that would refuse. Beyond mail and a connected calendar there is nothing: no contacts list past resolving an address from your own history, no files on a drive, no tasks.
-
Watch the mailbox and act on its own
Nothing here runs unless a client calls it. There is no rule engine and no background process reading your mail. An assistant does something because somebody, or a workflow somebody built, asked it to.
-
Destroy a message
Delete moves to Trash and stops there. No tool empties a folder and no tool erases a message, which means nothing on this surface can lose your mail for good.
-
Read a mailbox you have not connected
Every tool is scoped to the one mailbox a connector was signed in for. An assistant holding two of your mailboxes holds two separate connectors, and neither can see the other.
Questions
Questions people ask before trying any of this
How many MCP calls does a normal day use?
Somebody using this conversationally rather than in bulk tends to spend a handful of calls per session: a search, a couple of reads, a reply. The free tier of 5 calls a day per mailbox covers trying any single use case on this page. A daily triage habit or an unattended agent runs into it quickly, which is what the Pro ceiling is for.
Are the call counts on this page exact?
No, and they say "about" for that reason. How many calls a request costs depends on how the model chooses to work: one assistant answers a search question in one call, another reads three of the results before summarising. The ranges are what we see rather than a specification, and the control panel shows your own real figures per mailbox.
Does asking for a summary mark my email as read?
No. Reading through this server never sets the seen flag, on any tool. Marking something read is a separate tool that runs only when you ask for it, and it is reversible. This is checked against a real mailbox rather than asserted, and it is the one behaviour on this page that no other vendor documents either way about its own product.
Can it attach a file from my computer?
Yes. The assistant asks for a one-off upload link, and either sends the file to that link itself if it can run commands, or hands you the link to drop files on. What arrives lands in your own Drafts folder as a draft you can open, keep or delete, so the files are in your mailbox rather than on our servers, and the link stops working after half an hour. The assistant then attaches them the same way it attaches anything else already in the mailbox: fetched at send time, never passing through the model, inside the ordinary 10 MB limit on the whole message. A file already in your mailbox and a file at a web address both work without any of that. Only something the assistant composes itself has to be far smaller, a few tens of kilobytes, because the model has to produce those bytes as output.
Can it delete my email permanently?
No. Deleting moves a message to Trash, exactly as pressing Delete in Outlook would, and it is recoverable from there. There is deliberately no tool that empties a folder and no tool that destroys a message you have received or sent. One tool does remove something outright, and it is worth naming rather than hiding behind the word "nothing": updating a draft replaces it, so the version you changed your mind about is gone instead of sitting in your bin. The new version is saved first, and no message anybody sent you is ever removed this way.
Will my assistant do these things without asking me?
That is your client's decision rather than ours. Every tool is annotated: the read ones carry a read-only hint, and the ones that send, move or delete are marked so a host can stop and ask first. Most clients do ask before a destructive tool. What we control is the annotation; what your client does with it is in its own settings.
Does it work with a mailbox that is not Gmail or Microsoft 365?
Yes, and that is most of the point. Anything speaking IMAP and SMTP over TLS connects: your own domain, a web host's mailbox, Fastmail, Zoho, a personal outlook.com address. No native AI connector reaches any of those, which is set out in full on the native connectors comparison.
Which of these need the paid plan?
None of them technically. Every tool is on every plan, including the free one; what differs is how many calls a day you get. So the free tier is enough to try anything on this page against your own mail, and the paid one is what you move to when you start doing it every morning.
Pick one and try it on your own mail.
5 MCP calls a day on any mailbox, with no card. Enough for any single thing on this page, against real messages rather than a demo.