Inboxpilot-ai

mail monitoring agent

Comment

InboxPilot AI

AI-Powered Inbox Triage Agent

InboxPilot AI is an intelligent email triage agent built with the Strands Agents SDK. It analyzes unread emails, checks sender history, and decides whether each message should be drafted, escalated, or automatically handled.

Built for the "Agents for Humans" hackathon — Everyday / Professional Agents track.

How It Works

                    Email Inbox
                         |
                         v
                  InboxPilot AI
                         |
                         v
               Check Sender History
                         |
                         v
                  AI Classification
                         |
          +--------------+--------------+
          |              |              |
          v              v              v
      AUTO-REPLY       DRAFT         ESCALATE
          |              |              |
          v              v              v
    Routine emails   Human review    User decision

Actions

  • AUTO-REPLY — Routine, low-risk messages where an automatic response is appropriate.
  • DRAFT — Messages that need a human voice but do not require an important decision.
  • ESCALATE — Emails involving money, contracts, legal matters, or important personal decisions.

Key Features

  • 🤖 Local AI-powered email classification
  • 🧠 Sender history and memory
  • ✍️ Automatic draft generation
  • 🚨 Human-in-the-loop escalation
  • 🛡️ Code-level safety guardrails
  • 📋 Auditable email actions
  • 💻 Runs locally using Ollama
  • 🔌 Designed for future Gmail/Outlook MCP integration

Technology Stack

  • Python
  • Strands Agents SDK
  • Ollama
  • Llama 3.2
  • Strands Tools
  • JSON-based sender memory

The current demo runs locally using Ollama and does not require an Anthropic API key or AWS credentials.

Project Structure

inbox_triage_agent/
│
├── agent.py
│   └── Agent definition and email processing workflow
│
├── tools.py
│   └── Email actions and sender memory
│
├── mock_inbox.py
│   └── Six sample emails for demonstration
│
├── sender_memory.json
│   └── Sender handling history
│
├── test_ollama.py
│   └── Ollama and Strands tool-calling test
│
├── requirements.txt
│
└── README.md

Workflow

1. Find unread emails
        ↓
2. Select one email
        ↓
3. Check sender history
        ↓
4. Analyze email using Llama 3.2
        ↓
5. Select an action
        ↓
6. Apply safety checks
        ↓
7. Execute the appropriate action
        ↓
8. Update email status
        ↓
9. Process the next email
        ↓
10. Display final summary

Python controls the processing loop while the AI performs the email classification and generates response text.

Requirements

  • Python 3.13
  • Ollama
  • Llama 3.2
  • Strands Agents SDK

Installation

1. Clone the repository

git clone https://github.com/YOUR_USERNAME/inboxpilot-ai.git
cd inboxpilot-ai

2. Create a virtual environment

Windows CMD:

python -m venv venv

Activate it:

venv\Scripts\activate

3. Install dependencies

pip install -r requirements.txt

Install Ollama support:

pip install "strands-agents[ollama]"

Ollama Setup

Install Ollama:

https://ollama.com/download/windows

Check installation:

ollama --version

Download Llama 3.2:

ollama pull llama3.2

Verify:

ollama list

You should see:

llama3.2:latest

Ollama normally runs locally at:

http://localhost:11434

Run the Agent

Activate the virtual environment:

venv\Scripts\activate

Run:

python agent.py

The demo processes six sample emails from the mock inbox.

Example

============================================================
INBOX TRIAGE AGENT
============================================================

Found 6 unread emails.

Processing: e1 - Your package is out for delivery
  Decision: DRAFT
  ✓ Saved a draft reply ...

Processing: e3 - Can you review the Q3 proposal by Friday?
  Decision: DRAFT
  ✓ Saved a draft reply ...

Processing: e4 - Updated contract terms require your signature
  Decision: ESCALATE
  ✓ Escalated ...

Processing: e5 - Your bill is ready
  ⚠ Money-related email → escalating instead.
  ✓ Escalated ...

Final Summary

============================================================
INBOX SUMMARY
============================================================

[DRAFTED] Can you review the Q3 proposal by Friday?
  → draft ready for you: ...

[ESCALATED] Updated contract terms require your signature
  → needs your decision: ...

[ESCALATED] Your bill is ready
  → needs your decision: ...

Safety Guardrails

InboxPilot AI follows a human-in-the-loop approach.

Low Risk
   ↓
AI handles

Medium Risk
   ↓
AI creates draft
   ↓
Human reviews

High Risk
   ↓
AI escalates
   ↓
Human decides

Money Protection

The Python layer checks for money-related content before allowing an automatic reply.

Examples include:

bill
payment
money
price
cost
invoice
refund
billing
$

If detected, the email is escalated even if the AI incorrectly selects AUTO_REPLY.

AI Decision
     ↓
Safety Check
     ↓
Money-related?
   /       \
 Yes       No
  ↓         ↓
Escalate   Execute

Why InboxPilot AI?

People receive many emails that do not actually require their attention.

InboxPilot AI focuses on reducing unnecessary human involvement while keeping humans in control of important decisions.

100 Emails
     ↓
InboxPilot AI
     ↓
Routine → Handled
     ↓
Needs Voice → Drafted
     ↓
Needs Decision → Escalated
     ↓
Human

The goal is not to automate everything.

The goal is to automate the routine work and surface only what genuinely needs the human.

Current Demo

The mock inbox contains:

1. Delivery notification
2. Newsletter
3. Work proposal
4. Contract update
5. Electricity bill
6. Personal dinner message

This allows the complete agent workflow to be demonstrated without connecting a real email account.

Future Extensions

Real Email Integration

Replace the mock inbox with a Gmail or Outlook MCP connector.

Gmail / Outlook
      ↓
MCP Connector
      ↓
InboxPilot AI
      ↓
AI Classification
      ↓
Action

Notifications

Connect escalation to:

  • Slack
  • Microsoft Teams
  • Push notifications
  • Email notifications
  • Daily digest

Smarter Memory

Expand sender memory to track previous interactions and explain why a sender is trusted.

Stronger Guardrails

Add code-level restrictions for:

  • Financial messages
  • Legal messages
  • Contracts
  • Security requests
  • Sensitive personal requests

Scheduled Processing

Run the agent periodically:

Every 15 minutes
       ↓
Check inbox
       ↓
Triage new emails
       ↓
Surface only important decisions

Project Status

Completed

  • [x] Local Llama 3.2 integration
  • [x] Strands Agents integration
  • [x] Email classification
  • [x] Sender history
  • [x] Draft generation
  • [x] Escalation
  • [x] Auto-reply capability
  • [x] Money-related safety guardrail
  • [x] Auditable email status
  • [x] Mock inbox demonstration
  • [x] Strands tool calling

Future

  • [ ] Gmail/Outlook MCP integration
  • [ ] Real email sending
  • [ ] Real draft creation
  • [ ] Slack/push notifications
  • [ ] Persistent agent memory
  • [ ] Scheduled inbox processing
  • [ ] Advanced safety policies

Built With

Share this project:

Updates

Submission history