Build an Email Triage Agent in 30 Minutes

Before You Start

Prerequisites:

• OpenClaw installed (15-minute installation guide)
• Gmail or Outlook account (we'll use Gmail in this guide)
• Basic comfort with terminal commands (copy-paste level)
• 10-15 emails in your inbox to test with

What This Agent Does

Your Email Triage Agent will:

• Read unread emails from your Gmail inbox
• Categorize each email into Urgent / Important / Routine / Spam using Claude AI
• Flag urgent messages that need immediate attention (e.g., from your boss, customer complaints, payment failures)
• Draft replies to routine emails (e.g., "When are you available?", "Can you send the invoice?")
• Send you a summary via Telegram with the breakdown and urgent items highlighted
• Run automatically every morning at 8am (scheduled via cron)

Step-by-Step Guide

cd ~/.openclaw
nano AGENTS.md

# Email Triage Agent

**Role:** Inbox manager and email categorization specialist

**Responsibilities:**
- Read all unread emails from Gmail inbox
- Categorize each email as Urgent / Important / Routine / Spam
- Flag urgent messages requiring immediate human attention
- Draft replies to routine emails (save as drafts, don't send)
- Send daily summary via Telegram with breakdown and urgent items

**Skills:**
- gmail_read_messages: Read unread emails
- gmail_create_draft: Create draft replies
- telegram_send_message: Send summary to user

**Triage Rules:**
- URGENT: From boss, customers with "urgent"/"asap", payment/billing issues, server/system alerts
- IMPORTANT: Project updates, meeting invites, client requests, team questions
- ROUTINE: Newsletters, automated notifications, "thanks" replies, low-priority questions
- SPAM: Marketing, cold outreach, suspicious links

**Schedule:** Every day at 8:00 AM (managed via cron)

**Model:** Claude Sonnet (fast + cheap for email processing)

openclaw auth gmail

openclaw test-skill gmail_read_messages --limit 5

mkdir -p ~/.openclaw/prompts
nano ~/.openclaw/prompts/email-triage.txt

You are an email triage specialist. Categorize this email into one of four categories:

**URGENT** - Requires immediate attention (within 1-2 hours)
Examples: Boss/CEO messages, customer complaints, payment failures, system outages, "ASAP" or "urgent" in subject

**IMPORTANT** - Requires attention today
Examples: Project updates, meeting invites, client questions, team collaboration

**ROUTINE** - Can wait 1-3 days
Examples: Newsletters, automated reports, "thanks" replies, low-priority questions

**SPAM** - Marketing, cold outreach, suspicious
Examples: Unsolicited sales emails, promotional content, phishing attempts

For each email, respond with:
1. Category (URGENT/IMPORTANT/ROUTINE/SPAM)
2. One-sentence reason
3. Suggested action (Reply/Archive/Flag/Delete)
4. If category is URGENT or IMPORTANT and email asks a question: draft a short reply

Be strict with URGENT. Only 5-10% of emails should be urgent.

nano ~/.openclaw/scripts/email-triage.sh

#!/bin/bash
# Email Triage Agent - Runs daily to categorize inbox

# Read the triage prompt
PROMPT=$(cat ~/.openclaw/prompts/email-triage.txt)

# Get unread emails (last 24 hours)
EMAILS=$(openclaw run-skill gmail_read_messages --unread-only --limit 50)

# Run Claude to triage each email
RESULTS=$(echo "$EMAILS" | openclaw chat --agent email-triage --system "$PROMPT")

# Count categories
URGENT=$(echo "$RESULTS" | grep -c "^URGENT")
IMPORTANT=$(echo "$RESULTS" | grep -c "^IMPORTANT")
ROUTINE=$(echo "$RESULTS" | grep -c "^ROUTINE")
SPAM=$(echo "$RESULTS" | grep -c "^SPAM")

# Build summary
SUMMARY="📬 Email Triage Summary ($(date '+%Y-%m-%d %H:%M'))

🔴 URGENT: $URGENT
🟡 IMPORTANT: $IMPORTANT
🟢 ROUTINE: $ROUTINE
⚫ SPAM: $SPAM

$RESULTS"

# Send summary via Telegram
openclaw run-skill telegram_send_message --text "$SUMMARY"

echo "Email triage complete. Summary sent to Telegram."

chmod +x ~/.openclaw/scripts/email-triage.sh

~/.openclaw/scripts/email-triage.sh

crontab -e

# Run email triage agent every day at 8:00 AM
0 8 * * * ~/.openclaw/scripts/email-triage.sh >> ~/.openclaw/logs/email-triage.log 2>&1

What You Just Built

📬 Email Triage Summary (2026-03-15 08:00)

🔴 URGENT: 3
🟡 IMPORTANT: 7
🟢 ROUTINE: 18
⚫ SPAM: 12

=== URGENT ===

1. [Customer Support] Site is down - users can't log in
   → ACTION: Reply immediately, escalate to dev team

2. [CEO] Need Q1 numbers for board meeting today
   → ACTION: Reply with dashboard link

3. [Stripe] Payment failed for invoice #2847
   → ACTION: Update payment method

=== IMPORTANT ===

4. [Project Manager] Design review meeting moved to 2pm
   → DRAFT REPLY: "Thanks for the update. I'll be there."

5. [Client] Can you send the final proposal by Friday?
   → DRAFT REPLY: "Yes, I'll have it to you by Thursday EOD."

[... 5 more IMPORTANT emails ...]

=== ROUTINE & SPAM ===
18 routine emails (newsletters, automated reports)
12 spam emails (cold outreach, promotional)
→ Archived automatically

Customization Ideas

🎯 Make It Smarter

• Learn from your behavior: Track which emails you respond to quickly vs ignore. Update triage rules based on patterns.
• Sender VIP list: Always mark emails from specific people (boss, top clients) as URGENT.
• Keyword triggers: Flag emails containing "invoice", "payment", "urgent", "asap", "deadline".
• Auto-archive spam: Instead of just flagging, automatically archive emails categorized as SPAM (after 7 days of testing).

⚡ Add More Actions

• Auto-send routine replies: For emails like "Thanks!" or "Confirmed", send the draft reply automatically (add --send flag).
• Create tasks from emails: When an email asks for action, create a Notion task or Todoist item automatically.
• Weekly digest: Send a summary email every Friday with stats (% spam, most common senders, time saved).
• Slack integration: Post urgent emails to a private Slack channel instead of Telegram.

🔄 Multi-Inbox Support

• Multiple Gmail accounts: Run separate triage agents for work@company.com and personal@gmail.com.
• Outlook support: Replace gmail_read_messages with outlook_read_messages (same script structure).
• Aggregate summary: Triage 3 inboxes and send one combined Telegram summary.

Time & Cost Breakdown

Next Steps

You've just built your first autonomous AI agent. Here's what to do next:

1. Let it run for 7 days - Monitor the Telegram summaries and adjust triage rules if needed.
2. Build agent #2 - Try the Meeting Notes Agent (covered in Module 1: Security First).
3. Learn multi-agent orchestration - The full course teaches you how to run 5-10 agents working together (Module 7).
4. Join the community - Share your setup in the OpenClaw Discord and see what others are building.

Troubleshooting

Agent isn't running at 8am

• Check your crontab: crontab -l
• Verify the script path is correct (use absolute paths, not ~)
• Check logs: tail -20 ~/.openclaw/logs/email-triage.log
• Test cron is working: crontab -e and add a test job that runs in 2 minutes

Gmail authentication fails

• Make sure you selected the correct Gmail account during OAuth
• Check that "Less secure app access" is NOT enabled (we use OAuth, not passwords)
• Re-run: openclaw auth gmail --reset
• Verify credentials file exists: ls -la ~/.openclaw/credentials/gmail.json

Telegram messages not sending

• Authenticate Telegram: openclaw auth telegram
• Test the skill: openclaw run-skill telegram_send_message --text "Test"
• Check bot token is valid: openclaw doctor

Claude categorizing everything as URGENT

• Strengthen the prompt: Add "Only 5-10% of emails should be URGENT" at the end
• Give examples: Add 2-3 example emails with expected categories to the prompt
• Use temperature: Add --temperature 0.3 to the openclaw chat command (makes Claude more consistent)

Frequently Asked Questions

Is this safe? Can the AI send emails without my permission?

No. By default, OpenClaw uses read-only Gmail access. It can read emails and create drafts, but it cannot send emails unless you explicitly add the --send flag to gmail_create_draft. The script above only creates drafts - you review and send them manually.

What if Claude makes a mistake and marks something urgent as spam?

You'll still see all categorizations in the Telegram summary. Even if something is marked as spam, you can review it. After 7 days of testing and refining your prompt, the accuracy is typically 95%+. You can also add a "confidence score" to the prompt output.

Can I run this on multiple email accounts?

Yes. Create separate agent configurations in AGENTS.md for each account (e.g., email-triage-work and email-triage-personal), authenticate each account separately, and schedule separate cron jobs.

How much does this cost to run?

If you process 50 emails/day with Claude Sonnet: £3-9/month in API costs. Compare to Zapier (£30+/month) or Make.com (£29+/month). OpenClaw is 60-90% cheaper because you only pay for what you use.

What if I want it to auto-reply to some emails?

Modify the script to add --send flag to gmail_create_draft for emails that match certain criteria (e.g., subject contains "meeting confirmed" or sender is a specific person). Start with drafts-only for 14 days to build confidence.