Claude Gmail connectorWhat Claude can do in your Gmail account.
The Claude Gmail connector exposes 29 tools. 4 of them read your mailbox, 12 change something in it, and 13 are documented nowhere. This page covers what Claude sees, what it can send, what it asks you first, and where the official sources disagree.
Verified Trustpilot reviews · AI, automation & growth agency
You ask the question, Claude goes into the mailbox
Once the connector is plugged in, you stop pasting emails into the chat. You describe what you are looking for, Claude queries the mailbox, reads the threads that matter and answers with what it found. It runs on your own permissions: anything you cannot reach in Gmail yourself stays out of reach for it too.
Finding, first. You describe the message from memory, search_threads goes after it, get_thread pulls the full conversation back and get_message opens the exact email. No Gmail search operators to remember, just a question.
Preparing, next. create_draft parks an answer as a draft in your account, update_draft reworks it as many times as you need, and nothing leaves the mailbox until you say so. Sending runs through send_message, reply or forward, the three actions Claude checks with you every single time.
Filing, last. Seven label tools tag a message or a whole thread, create a label, take one off. Enough to clean up a mailbox where three clients share the same inbox.
What it will not do: open an attachment (it sees the file name, not what is inside), handle every advanced Gmail filter, and above all act on its own. Nothing fires when mail arrives. Claude calls these tools while it answers you, not after you close the tab.
That gap is where automation tools take over, and they do a different job: Gmail n8n integration and Gmail Make integration build scenarios that run without you, while this connector answers inside a conversation. The rest of the catalog sits in our Integrations hub.
Claude's vocabulary in one minute
Five words come back in every connector discussion. Here they are in plain English, once and for all.
- Connector
- The link you set up once between Claude and an account you already own, so it can work in that account while it answers you.
- Tool
- One named thing a connector lets Claude do in that account. Claude picks the ones it needs on its own, and the directory sheet lists them.
- Authorization
- The sign-in screen of the service itself, where you hand over the access Claude will use. Granted once per person, taken back the same way.
- Approval
- The confirmation Claude waits for before it does something that changes the account. It shows up in the chat, at the moment it matters.
- Directory
- The official catalog where every connector has its own sheet: who publishes it, what it is for, and the named list of tools it exposes.
- MCP
- The shared standard connectors are built on, the thing that lets an assistant talk to an outside service. You will see the acronym around.
Plug Gmail into Claude in three moves
- 01
Open the connector list
In Claude's settings, go into Customize, then Connectors. The list shows what is already plugged in and what is not. Gmail is in there under its own name, published by Google.
- 02
Start the connection from its row
Hit Connect on the Gmail row. Claude never asks for your password: Google opens its own window, and you sign in there with the account you actually want to open up, that one and no other.
- 03
Read the authorization screen, then approve it
That last screen belongs to Google, not to Claude, and it is the one deciding what the access covers. Read it fully before you accept. A connection that broke gets fixed with Disconnect and a fresh Connect.
The 29 tools of the Claude Gmail connector, one by one
Gmail gives Claude 29 tools: 4 that read your account, 12 that change something in it, and 13 no official source describes.
They sit in three groups: what Claude reads, what it writes, and what no official source describes. The names are the ones Claude shows you, in English, and they never get translated.
- 4 read
- 12 write
- 13 not documented
What Claude reads (4)
4 toolsFour tools that look at the mailbox without changing anything in it.
get_message
One email, opened and handed back in plain words: who sent it, when, the subject line, the body. Claude works from that instead of asking you to paste it in.
get_thread
A full conversation comes back in one go, in order, with the replies stacked and everyone who took part in it. Claude gets the story of the exchange rather than one isolated quote.
list_drafts
Drafts already sitting in the account come back as a list, with enough detail to recognize them. Nothing gets rewritten at this stage: Claude is looking at what is waiting to be finished.
search_threads
Plain language is enough here: Claude turns your request into a search and brings back the conversations that match. No need to know Gmail's search operators or to build a query by hand.
What Claude writes (12)
12 toolsTwelve tools that change something in the account. Three of them have their confirmation spelled out by an official source. For the other nine, the general rule below is what applies.
create_draft
Approval: see the ruleClaude writes the message and saves it as a draft in the account. The text waits for you in Gmail, readable and editable before anything goes out.
create_label
Approval: see the ruleA new label appears in the mailbox's organization, ready to take messages. This is filing, not writing mail.
delete_label
Approval: see the ruleA label that outlived its purpose leaves the mailbox's organization. Among the seven filing tools, this is the one that removes rather than adds, so it pays to know exactly which label you are aiming at.
forward
Asks you firstA message you already received goes out to someone else, along with the thread it came from. It leaves your mailbox toward a third party, which is a different thing from an internal note.
label_message
Approval: see the ruleA message gets a label and lines up with the others carrying it. The move targets a single email, not the conversation around it.
label_thread
Approval: see the ruleThis time the label lands on the entire conversation rather than one message inside it. Handy when an exchange has to live in one place from first email to last.
reply
Asks you firstThe answer goes out inside the existing conversation, under the messages already exchanged. The recipient reads it as a reply coming from you.
send_message
Asks you firstA brand new message leaves the account, recipients and subject line included. This is the tool behind the question everybody asks first: yes, Claude can send an email on your behalf.
unlabel_message
Approval: see the ruleThe label sitting on one specific message is removed. The tool works on how the mailbox is filed, not on the mail itself.
unlabel_thread
Approval: see the ruleSame move, one level up: the label leaves the full thread in one block instead of message by message.
update_draft
Approval: see the ruleA draft already saved gets picked up and rewritten: the tone, the length, whatever was missing. The text stays a draft, it does not turn into a send.
update_label
Approval: see the ruleAn existing label gets modified. The sources talk about label management as a whole and never spell out what this update actually covers: the official sheet stops there.
Undocumented (13)
13 tools"Undocumented" means neither harmless nor hidden. The directory publishes the names of these thirteen tools and nothing else: no official source says what they do, and we refuse to guess it from a name.
apply_sensitive_message_label
The directory prints this name in the connector's tool list and stops there. No official page describes what Claude does with it, so this page will not describe it either: a tool name is not documentation.
apply_sensitive_thread_label
Same silence around this one: not the help center, not Google's documentation, not the connector sheet. Better to say so plainly than to fill the hole with a guess that reads well.
create_filter
Nothing published on this tool, anywhere. Its name appears in the official list and that is the whole of what the directory gives: saying more would mean guessing, which this page refuses to do.
list_filters
Not one official sentence sheds light on this one. The precise tool-by-tool labeling only exists in the settings of an account that is already connected, and it is published on no public page. A name does not stand in for it.
list_labels
This one shares the fate of the other twelve: in the list, absent from every explanation. Its name looks like a lookup tool, and that is exactly the shortcut this page declines to take.
mark_message_spam
No official documentation covers this tool. The directory publishes the name, the help center never picks it up, and the publisher says no more about it. This page flags it as undocumented rather than lending it a behavior.
mark_thread_spam
Not a single official line on this one. It shows up in the connector's tool list with no sentence attached: nothing about what it touches, nothing about what it would change, nothing about what it would ask you.
trash_message
Official sources say nothing about this tool, and its name changes none of that. Reading a capability into a name is precisely what produced the wrong pages floating around on this subject.
trash_thread
Still nothing documented here. The September 2026 sweep found no official sentence describing this tool, and this page would rather say so than invent a plausible explanation.
unmark_message_spam
The help center never mentions this tool, and the connector sheet does nothing but name it. Nobody publishes what Claude would do with it, so this page will not publish it either.
unmark_thread_spam
There is no official description for this one, no more than for its neighbors in the list. The directory sticks to the name, and this page sticks to the directory: nothing will be claimed about what it touches.
untrash_message
Neither the directory, nor the help center, nor the publisher's own documentation covers this tool. The official list names it, full stop. That is a gap in the source, not an oversight on this page.
untrash_thread
Last in the list, and last without documentation. No official source says what Claude would do with it, and thirteen tools out of twenty-nine are in that position on this connector.
What Claude asks before it acts
By default, Claude stops and asks for your go-ahead before every action it takes in your account on your behalf. The prompt shows up in the conversation, at the moment it counts.
On a Team or Enterprise workspace, owners decide whether a member can let certain actions run without being asked each time. They can also fence in what the connector is allowed to do for the whole organization, leaving reading open and closing writing, for instance. That setting applies to everybody and nobody works around it from a personal account. Outside those workspaces the rule is the same for everyone, and it is the one above: nothing is silently agreed to on your behalf.
Which plans this connector is available on
Across the 819 sheets in the directory, not one displays plan availability: the per-connector answer is published nowhere. That is a hole in the catalog, not a gap in our sweep.
The general rule is published, though: remote connectors are open to all users on Claude, on Cowork, on Claude Desktop and on mobile. On Team and Enterprise, an Owner or a Primary Owner opens the connector for the organization before members can each sign in to it. For the current state of this one, its sheet in the official directory is the only place worth checking.
What this connector will not do
A connector is not an automation. Its tools get called by Claude while it answers you, so nothing fires when mail arrives and no mailbox sorts itself overnight while you sleep.
A tool list is an observed floor, never a guarantee: an administrator can open write access that appears on no public sheet. The directory's verification badge is not a security audit either. Anthropic prints it on every sheet: it does not decide which tools a publisher exposes and does not guarantee they behave as advertised. So you only plug in connectors whose publisher you know, and here that publisher is Google.
Can Claude actually send an email from Gmail?
- Developer documentation ↗September 2026Out of dateClaude can neither create, nor send, nor change an email.
- Help center ↗August 2026Claude sends, replies to and forwards mail out of Gmail.
- Directory sheet ↗September 2026send_message, reply, forward and create_draft all appear in the connector's public tool list.
What this page followsYes, it sends, and it asks you first. The help center and the directory sheet are the ones we follow; the developer documentation describes a state of things from before August 2026 and we set it aside. Pages still claiming the opposite copied the outdated version.
Need help connecting Gmail to Claude?
A person reads every message.
Frequently asked questions
01What can the Claude Gmail connector do?
02Can Claude write, send or delete emails?
03Does Claude ask before it acts?
04Which plans is this connector available on?
05Does Claude see my whole Gmail account?
06Does Claude read the attachments in my emails?
07Claude or an automation tool for Gmail?
Get our weekly automation tips.
No spam. Unsubscribe anytime.