- Home
- Skills
- Productivity
- Inbox Triage: Gmail Sorter with Approval Step
More screenshots
Works with the AI Agents you already use
Inbox Triage: Gmail Sorter with Approval Step
**inbox-triage** reads your recent unread Gmail through a Gmail MCP connector, groups messages that share a subject, and proposes a bucket for each group: Reply, Later or Noise.
Free
Inbox Triage: Gmail Sorter with Approval Step
Example session with this skill installed
Triage my inbox.
- Read your context and instructions
- Compiled the inbox triage
Illustrative sample. Every sender, subject and link below is fictional and was written to show the format. It is not output from a real inbox.
Unread in the last 24 hours: 12 messages in 6 groups.
| Group | Count | Bucket | Reason | Newest message |
|---|---|---|---|---|
| Invoice due Friday | 1 | Later | Subject contains "invoice", so it is never Noise. Check the due date. | Billing team |
| Can you review the proposal? | 3 | Reply | Dana (client) asks you a direct question. Two teammates replied on the same subject. | Dana (client) |
| Invitation: Planning sync | 1 | Later | Calendar invitation, so it is never Noise. Accept or decline from your calendar. | Calendar |
| Weekly Dev digest | 1 | Later | Newsletter you open. Read when you have time. | Newsletter: Weekly Dev |
| Build passed | 4 | Noise | Automated notification. The same message repeated four times. | Build bot |
| Weekend sale | 2 | Noise | Marketing mail from two senders with the same subject. | Shop promotions |
Reply: 1 group. Later: 3 groups. Noise: 2 groups (6 messages).
Nothing has been changed in your mailbox. Mark the Noise groups as read?
Connects securely to your tools. The creator never sees your data.
About this skill
inbox-triage reads your recent unread Gmail through a Gmail MCP connector, groups messages that share a subject, and proposes a bucket for each group: Reply, Later or Noise. It then shows a table and stops. Only if you say yes does it mark the Noise groups as read.
How it works
- Fetches unread mail from the last 24 hours.
- Groups messages by normalized subject (strips "Re:", "Fwd:" and ticket numbers). The same subject from different senders is one group.
- Picks one bucket per group:
- Reply: asks you a question, assigns you a task, or comes from a person waiting on you.
- Later: informational but relevant, such as newsletters you open or threads you are cc'd on.
- Noise: automated notifications, marketing, or duplicates of something already in Reply or Later.
- Shows a table (group, count, bucket, one-line reason, link to the newest message) and asks: "Mark the Noise groups as read?"
- Only after a yes, marks every message in the Noise groups as read and reports the count.
Guardrails
- Never-noise list. These always go to Reply or Later, whatever else the message says: subject or body containing invoice, contract, payment, security alert or password; calendar invitations; mail from a sender you replied to in the last 30 days.
- The only write action is "mark as read" on messages you approved.
What it does not do
- It does not send, reply to, or forward mail.
- It does not archive or delete anything.
- It does not mark anything as read before you approve the table.
- It does not touch Reply or Later groups.
- It does not look further back than the last 24 hours.
- It does not include a Gmail connector. You supply your own.
Setup
Needs Claude Code and a Gmail MCP connector. The included settings.example.json allows read and label tools and denies send, reply, forward and trash. Its tool names follow the mcp__gmail__* pattern, so rename them to match your connector. Agensi delivers the Skill and its text files through MCP, and it does not change your Claude Code settings. The permission rules in settings.example.json are a reference, so copy them into your own settings by hand.
The deny list is a Claude Code permission rule that matches on tool name. It is not a sandbox: a differently named tool is not covered. The never-noise list is written instructions in SKILL.md, not code, so edit it for your own mail before the first run.
Watch it built
EP02 on the Guardrail Dev channel: https://www.youtube.com/watch?v=U24Jbg2_sdI (the inboxes shown on screen are sample data, not a real inbox)
Files
Source repo: https://github.com/guardraildev/inbox-triage-skill
More Claude Code guardrail tutorials: https://www.youtube.com/@GuardrailDev
Gmail only, through a Gmail MCP connector you supply; the example tool names (mcp__gmail__*) may need renaming. Looks only at unread mail from the last 24 hours. The buckets and the never-noise list are written instructions the model follows, not enforced code, so read the table before you approve. The deny rules in settings.example.json match tool names and are not a sandbox, and Agensi does not install them: copy them into your Claude Code settings by hand. It never sends, replies, forwards, archives or deletes.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 13 days ago
- Free with an Agensi account
Requires
Permissions
Allowed Hosts
The Skill itself runs no scripts and makes no network calls of its own. Network access is the Gmail MCP connector the user supplies. The only write is marking messages as read, and only for Noise groups the user approved. No environment variables, secrets or local files are read. Terminal is needed only for the optional cron example in the repo README, which is not in the zip. If the Agensi scan flags anything, the likely cause is the words send, reply, forward and trash in settings.example.json, where they are deny rules.