manor-os/manor-ai/mcp-gmail

Gmail Runtime Skill

Use this skill to operate the user's connected Gmail account through the Gmail MCP (mcp__gmail__*). Do not use web_search, web_fetch, or a browser to read or send mail when this account is connected — these tools are the source of truth for the user's inbox.

When To Use

Use Gmail when the user names Gmail / Google Mail, or asks to read, search, summarize, triage, send, reply, draft, label, or archive email and a Gmail integration is connected. For non-Gmail IMAP/SMTP accounts use the email MCP; for Outlook use the outlook MCP — the tool surfaces differ, do not mix them.

Connection

If a tool returns an auth error (token expired, insufficient scope, account not connected), stop and tell the user to reconnect Gmail in Integrations. Do not fall back to scraping mail through a browser. mcp__gmail__get_profile is a cheap way to confirm the connected address before acting on "my email".

Core Tools

Read / search:

• list_messages — search with Gmail query syntax (required: query, e.g. is:unread newer_than:7d, from:x@y.com). Returns IDs only.
• get_message — full headers + body by message_id.
• list_threads / get_thread — work at thread granularity for conversations.
• download_attachment — fetch an attachment by message + attachment id.
• list_labels / get_profile — labels and the connected address.

Write / send (high-impact — see Guardrails):

• send_message — required: to, subject, body; optional cc, bcc, reply_to_message_id.
• reply_to_message — required: message_id, body (replies in-thread).
• list_drafts / get_draft, create_draft / update_draft / delete_draft, then send_draft — inspect and stage mail before sending; preferred when the user should review first.

Triage:

• mark_read / mark_unread, archive_message, trash_message (reversible ~30d via untrash_message), mark_spam.
• trash_thread — move a whole conversation to Trash after confirming the thread.
• batch_modify — add/remove labels across up to 1000 message_ids in one call; use for bulk triage instead of looping single calls.
• create_label / delete_label, add_label / remove_label — manage custom labels and apply them to messages.

Common Recipes

Summarize unread inbox

1. list_messages with query: "is:unread in:inbox newer_than:7d".
2. get_message (format metadata or full) for each returned ID.
3. Summarize; do not mark read unless the user asked.

Reply to a specific email

1. list_messages to locate it (e.g. from:... subject:...), then get_message to read full context.
2. Draft the reply text, show it to the user, and get confirmation (see Guardrails).
3. reply_to_message with that message_id and body — keeps the thread.

Bulk triage / clean up

1. list_messages with a precise query (e.g. from:newsletter@x.com older_than:30d).
2. Confirm the count and the action with the user.
3. batch_modify (e.g. add TRASH / remove INBOX) over the collected IDs — one call.

Guardrails

• Never send or reply without explicit user confirmation of the recipient(s) and the body. Default to create_draft and show the draft when intent is ambiguous; only send_message / send_draft after the user approves.
• Verify recipients before sending — re-read to/cc/bcc; do not invent addresses. Use bcc for multi-recipient external sends to avoid leaking address lists.
• Bulk operations (batch_modify, mass trash) require an explicit count + confirmation. Prefer archive_message/trash_message (reversible) over mark_spam or hard deletes.
• Privacy: only read messages relevant to the task. Do not dump full inbox contents into other tools or external channels.

Edge Cases & Errors

• list_messages returns IDs only — you must get_message to read content; don't assume the query result has bodies.
• Empty result ≠ error: report "no matching messages" rather than retrying with broader queries unprompted.
• Pagination: use the returned page_token to continue; don't raise max_results past 100.
• trash_message is reversible (~30 days) via untrash_message; there is no separate permanent-delete tool here — say so if the user expects hard deletion.
• Auth / scope errors → stop and ask the user to reconnect; do not work around with a browser.