Let agents work in Gmail without handing them the whole inbox
Email mixes routine retrieval with consequential actions. AIP turns Gmail's API into separate tools for reading, drafting, sending, labeling, and moving messages — so an inbox analyst and an outbound assistant never need the same authority.
What your agent can do
Gmail tools follow the operation names stored by AIP's OpenAPI discovery. GET operations are read tools; drafting, sending, labeling, modifying, and moving messages are writes. Each can be exposed independently to an agent, including separately connected inboxes.
| Tool | What it does | Access |
|---|---|---|
| gmail_users_getprofile | Read the connected mailbox profile and current message and thread totals. | read |
| gmail_users_messages_list | Search or list message IDs using Gmail query syntax, labels, and pagination. | read |
| gmail_users_messages_get | Fetch one message with its headers, body structure, labels, and thread reference. | read |
| gmail_users_messages_attachments_get | Retrieve an attachment payload referenced by a message part. | read |
| gmail_users_threads_list | Search or list conversation threads in the connected mailbox. | read |
| gmail_users_threads_get | Fetch a whole thread and the messages it contains. | read |
| gmail_users_labels_list | List system and user labels available in the mailbox. | read |
| gmail_users_drafts_list | List saved drafts without sending them. | read |
| gmail_users_drafts_get | Fetch one draft and its underlying message content. | read |
| gmail_users_drafts_create | Save a new draft message for later review or sending. | write |
| gmail_users_drafts_update | Replace the message stored in an existing draft. | write |
| gmail_users_drafts_send | Send a previously saved draft. | write |
| gmail_users_messages_send | Send a new MIME message through the connected Gmail account. | write |
| gmail_users_messages_modify | Add or remove labels on one message, including read and starred state. | write |
| gmail_users_messages_trash | Move a message to Trash without permanently deleting it. | write |
| gmail_users_messages_untrash | Restore a message from Trash. | write |
AIP discovers 29 curated Gmail operations from the current public spec. These highlights focus on everyday inbox work and omit operations that the connected OAuth scope cannot perform.
Capability snapshot: 1 September 2026 · provider source
Treat reading and sending as different jobs
An assistant that prepares a morning inbox brief only needs search, message, thread, and attachment reads. A correspondence assistant may need draft tools but still no direct send. A tightly supervised support agent can send replies and adjust labels, while trash remains hidden. Gmail exposes those as separate operations, and AIP keeps the distinction when the tools reach your client.
The control room is where that policy lives. It can also separate personal and work inboxes instead of pooling both identities behind one opaque connector. See the rest of the integration catalog when an email workflow needs context from a calendar, tracker, or knowledge base.
Connect a Google inbox
- In the AIP dashboard, add Gmail and sign in to the exact Google account the agent should use. Google shows the requested Gmail access before consent.
- Attach that account to an agent and review each discovered operation. A practical default is Allow for list/get, Confirm for drafts and labels, and Confirm or Hide for send and trash.
- Add the generated AIP endpoint to the client. MCP clients use it directly; clients that prefer OpenAPI can use the document AIP generates for the same scoped agent.
There is no Gmail password or app password to paste into an agent config. AIP stores the OAuth credential and refreshes it behind the endpoint. The MCP/OpenAPI explanation shows why the same account policy can travel between Claude, Codex, and other clients.
Build review into the prompt
Permissions are the hard boundary; prompt structure adds a useful human checkpoint. Ask the agent to return candidate messages or a draft first, then approve a second step that changes the mailbox. This makes false-positive searches cheap to correct before a label, send, or trash call runs.
If several inboxes or specialized agents are part of the workflow, the pricing page describes the available account and agent capacity. Keep sensitive inboxes on their own connection so their tools can be hidden from agents that never need email context.
What to ask your agent
Inbox prompts are safest when they separate analysis from action. Ask for a reviewable result first, then make the send, label, or trash step explicit: