Donezo Inbox connector for Claude and Codex: reference

Updated October 19, 2026. Written from the connector's source code; setup commands checked against Anthropic's and OpenAI's documentation on October 19, 2026.

The connector is an MCP server over HTTP that gives an AI agent 22 tools for every Gmail and Google Workspace inbox you have added to Donezo Inbox. You create a private link in Settings, choose what it may do, and add it to Claude, Claude Code or Codex. This page is the reference; the how-to guide is the short version.

How do you install it?

In Donezo Inbox, open Settings, then Claude & Codex.
Choose what the link may do (see the table below), click Create agent link, and copy the link. It is shown once.
Add it to your agent with one of the commands below, replacing <link> with the link.

Claude Code (all your projects)

claude mcp add --transport http --scope user donezo <link>
claude mcp list
# to remove it later:
claude mcp remove donezo

Codex

codex mcp add donezo --url <link>
codex mcp list

Codex, in ~/.codex/config.toml instead

[mcp_servers.donezo]
url = "<link>"

Claude on the web and desktop: Customize, then Connectors, then + Add, then Add custom connector; name it, paste the link as the server URL and click Add. Turn it on in a conversation from +, then Connectors. On Team and Enterprise plans an Owner adds custom connectors in Organization settings. Anthropic allows one custom connector on its Free plan.

The link carries its own key, so treat it like a password: anyone with it can use it. Revoke it in the same settings page, and create a new one, if it leaks. Setup steps checked against Claude Code's MCP documentation, Anthropic's custom connector guide and OpenAI's Codex MCP documentation.

What can each link do?

Each link has one of four settings, chosen when you create it. The default is Read and write, with approval.

Link settings and what happens to each kind of call
Setting	Read and search	Drafts	Organize (archive, snooze, labels, reminders)	Send, schedule, forward, delete, rules, invitations
Read only	Runs	Refused	Refused	Refused
Drafts only	Runs	Runs	Refused	Waits for approval
Read and write, with approval (default)	Runs	Runs	Runs	Waits for approval
Read and write	Runs	Runs	Runs	Runs, with no question

Read only: List, read, search and ask. Cannot change anything.
Drafts only: Also saves drafts. Sending, scheduling, forwarding and deleting wait for your OK.
Read and write, with approval: Also archives, snoozes, labels and sets reminders. Sending, scheduling, forwarding and deleting wait for your OK.
Read and write: Everything, with no questions asked.

Separately, the Let agents send email switch (on by default) decides whether agents may send at all. Off, agents and scripts using an API key can only create drafts, whatever the link allows.

How do approvals work?

A call that needs approval is saved with a preview (account, recipients, subject and the start of the message, or what will happen) and shows under Settings, then Developer, then Approvals. Nothing happens until you approve it.
A request waits 24 hours, then expires. One link can have at most 25 open requests and one account 100, so an agent in a loop fills its own queue, not yours.
Moving mail to Trash or Spam counts as a delete, however it is asked for, including adding a Trash label.
Every call is written to an activity log with the tool, the time and a short summary, never the mail. The log is kept for 90 days.
Reading, searching and drafting work on every plan. Calls that change mail need Pro after 30 free actions on a free account.

Which tools does the connector have?

22 tools. Most take an optional account; leave it out and the call runs across every linked inbox, with each result tagged by account. Search uses Gmail's own syntax.

Every tool, with its class under the default setting
Tool	What it does	Class
`list_accounts`	Every linked inbox with its inbox and unread counts.	Read
`inbox_overview`	Mail from people first, then a count of everything else, across inboxes.	Read
`list_threads`	Gmail search across one or all inboxes, newest first (up to 50 per inbox; default query in:inbox).	Read
`get_thread`	A whole conversation: senders, recipients, dates, text and attachment names.	Read
`list_drafts`	Drafts in one or all inboxes.	Read
`list_labels`	The labels in an inbox.	Read
`list_snippets`	Saved snippets (reusable templates).	Read
`reply_zero`	What needs your reply, and what you sent that nobody has answered in two days or more.	Read
`list_rules`	Your plain-English inbox rules.	Read
`assistant_activity`	What the inbox assistant did recently.	Read
`meeting_briefs`	Meetings in the next 36 hours with who each person is and what you last emailed about (needs calendar access).	Read
`who_is`	Recent conversations with a person, across inboxes.	Read
`create_draft`	A new email or reply saved as a Gmail draft.	Draft
`send_draft`	Sends an existing draft.	Needs approval
`send_email`	Sends now, or schedules with sendAt (an ISO time, or a wall clock plus timeZone). Optional follow-up.	Needs approval
`update_threads`	Done (archive), back to inbox, read, unread, star, labels; or trash and spam. Up to 100 conversations per call.	Organize, or needs approval
`snooze_thread`	Out of the inbox until a time; it comes back unread at the top.	Organize
`set_reminder`	Brings a conversation back at a time if nobody has replied.	Organize
`set_followup`	Brings a conversation back after 1 to 30 days (default 3) if nobody replies, or cancels that.	Organize
`add_rule`	A plain-English rule the assistant applies to new mail.	Needs approval
`delete_rule`	Removes a rule by id.	Needs approval
`create_event`	A Google Calendar event, with Meet, Zoom or Teams. Not for changing or deleting events.	Organize, or needs approval

update_threads is an organizing call unless it trashes or marks spam, which needs approval. create_event is organizing unless it invites guests or rooms, which follows the same rule as sending.

What do example requests look like?

Each example names the tools it uses and only asks for what the tools accept.

Morning triage (inbox_overview, reply_zero; read only)

Using Donezo, list what needs my reply today across all my inboxes, then what I am still waiting on. Do not change anything.

A reply with context (list_threads, get_thread, who_is, create_draft)

Find the latest thread from priya@example.com, check our recent conversations, and save a draft reply proposing Thursday at 2pm. Do not send it.

A scheduled send (send_email with sendAt and timeZone; waits for approval by default)

Schedule that draft's text as a reply in the same thread for 2026-10-20T09:00 in America/Los_Angeles. Tell me the exact time before you queue it. (The draft itself stays in Drafts; send_email schedules a message, not a saved draft.)

Chase what is stuck (reply_zero, set_followup)

Which emails I sent have had no reply for more than three days? Set a follow-up in 2 days on each one from a person, not a company, and list them.

What does the connector not do?

It works with Gmail and Google Workspace accounts only.
It does not delete permanently: trash is the furthest it goes, and Gmail empties Trash after 30 days.
It cannot change or delete calendar events, only create them.
Conversations with people you have excluded from AI are withheld from agents.
Mail text is returned marked as untrusted, so the agent treats instructions inside an email as content, not commands. Outgoing text is checked before anything is saved or queued.
A very large request across many inboxes may run only as many as fit in one call; the reply names the inboxes it skipped.

Questions

Does the connector work without Donezo Inbox open?

Yes. It runs on Donezo's server, so the agent can use it while the app is closed. Scheduled sends are sent from the server too.

Can I have more than one link?

Yes. Create one link per agent, each with its own setting, so you can revoke one without the others.

Is there an API as well as MCP?

Yes. The same permissions apply to API keys, which are created under Settings, then Developer.

What does Donezo record about agent calls?

The activity log keeps the tool, time and a short summary for 90 days. Product analytics records which tool was called and whether it failed, never the arguments or any mail.