MCP tools overview
Every MCP tool inboxmcp exposes — read, search, draft, move, send — and which plan unlocks which.
What Claude / ChatGPT / any other MCP client can do through inboxmcp. Which tools are available depends on your plan.
Read tools (all plans)
| Tool | Description |
|---|---|
list_mailboxes |
Mailboxes available on the account |
list_folders |
Folder tree of a mailbox |
search_mails |
Search per folder (full text, sender, date, flags) |
read_mail |
Full message body and attachments |
read_mail_chunk |
Read a very long message in parts |
get_attachment |
Deliver an attachment via a short-lived signed URL |
list_capabilities |
Which tools the current plan unlocks |
Write tools (Premium / Business)
| Tool | Description |
|---|---|
draft_mail |
Create a draft (reply / forward aware, auto-quotes the original) |
edit_draft |
Revise an existing draft |
create_folder |
Create a new folder |
move_mails |
Move messages to another folder, the trash included |
send_mail |
Send a mail — the MCP client asks for confirmation before dispatch |
Not there yet: marking messages read, unread or starred (flag_mail). It is on the
roadmap. A separate delete_mail is not planned — to delete, the AI moves
the message to the trash with move_mails, where you can still get it back.
Security — what’s inside read_mail
Mail bodies returned from read_mail and read_mail_chunk are wrapped in
<untrusted_mail_body>…</untrusted_mail_body> markers. The AI assistant sees a
clean boundary between instructions (your chat input) and data (someone else’s
mail content), so a directive smuggled into a mail body — the classic “ignore
all previous instructions” — doesn’t get followed as an instruction. Details at
Security & privacy.
Every tool call lands in the audit log — Business accounts can view it in the portal.