Offmail: Local Email Agent

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend.
 

Why I made Offmail and the Story behind It

Arpit is one of my best friends and recently he started sending a lot of LinkedIn connection requests to recruiters, founders, and people he wants to learn from and make his online presence on LinkedIn
 
The problem wasn't getting connections accepted. It is what happens next.
A recruiter accepts the request. A founder replies. A potential opportunity appears. He sees the notification, thinks:

"I'll reply later."

And then later never comes. Days pass, the moment is gone, and another valuable connection turns into a missed opportunity. At first, I suggested using AI agents and automation tools. In theory, they could help manage replies and follow-ups.

But there were a few practical problems:

• Arpit comes from an Economics background, not a technical one.
• Most AI agent solutions require complicated setup and configuration.
• Many tools require connecting sensitive personal accounts and emails.
• Privacy was a major concern for him.
• A lot of existing solutions come with recurring subscription costs.

The more we talked about it, the more I realized the problem wasn't a lack of technology. The problem was that the available technology wasn't accessible to people like him.

I had the idea of building something simple and privacy-conscious that anyone could use online. However, between classes, internships, and studies, I rarely found enough uninterrupted time to work on personal projects. Like many side projects, the idea stayed in my notes for quite a while.

Then came Hacktoberfest.
For the first time, I had both a reason and an opportunity to finally build it. So I started working on Offmail.

The goal was simple:
Help Arpit respond to important conversations before opportunities slip away.

I built the first version, installed it on Arpit's computer, and watched him use it in real situations. The product is still in its early stages. There are plenty of features to build, improvements to make, and lessons to learn.

But seeing the person who inspired the idea genuinely happy using something I created was enough to convince me that this problem is worth solving.

And that's how Offmail started.

Demo

Here is a quick demo of Offmail in action:

Most importantly:The inbox never leaves the user's machine.There is no OffMail cloud. There is no AI API. There is no server processing private emails. Everything runs locally.

Source Code and Documentation

Qisanxi / Offmail

𝗹𝗼𝗰𝗮𝗹-𝗳𝗶𝗿𝘀𝘁 𝗔𝗜 𝗲𝗺𝗮𝗶𝗹 𝗮𝗴𝗲𝗻𝘁 𝘁𝗵𝗮𝘁 𝗵𝗲𝗹𝗽𝘀 𝘁𝘂𝗿𝗻 𝗟𝗶𝗻𝗸𝗲𝗱𝗜𝗻 𝗰𝗼𝗻𝗻𝗲𝗰𝘁𝗶𝗼𝗻 𝗻𝗼𝘁𝗶𝗳𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀 𝗶𝗻𝘁𝗼 𝗽𝗲𝗿𝘀𝗼𝗻𝗮𝗹𝗶𝘇𝗲𝗱 𝗳𝗼𝗹𝗹𝗼𝘄-𝘂𝗽𝘀.

Offmail

Offmail reads your Gmail's Social tab locally, detects LinkedIn "connection accepted" emails, drafts a personalized reply using Gemma 3 1B running on your laptop via Ollama, and you hit send — routed via LinkedIn's reply-to email address so it lands as a LinkedIn DM. Your inbox never leaves your machine.

Submitted for the Hacktoberfest 2026 Weekend DEV Challenge — theme: Build for a Friend.

---

🎥 Demo Video

Video Link : (https://youtu.be/xfp9BkdUpX0?si=6PghaHGxsBLSOUQ0)

---

---

Live Links

**URL : (https://offmail-seven.vercel.app/#get)

Why Offmail

Closed inbox-AI tools (Superhuman, Shortwave, etc.) read your emails on their servers. That means:

• Your private recruiter conversations sit in someone else's database
• They cost $20–30/month forever
• They stop working the moment you stop paying
• You can't audit how your data is used

Offmail flips that:

Hacktoberfest prompt question	Offmail's answer
Does it run on a laptop with no internet?	Yes — Ollama + Gemma 3

…

View on GitHub

How I Built It

Offmail is designed around a local-first architecture. Here are the core technologies and how they fit together:

• Open-Source AI: Google's Gemma 3 1B (an incredibly fast, open-weight model) running on Ollama for 100% local inference.
• Backend: FastAPI (Python) handles the API, IMAP/SMTP operations, and the local SQLite database.
• Frontend: A React UI powered by Vite 8 and Tailwind CSS v4, running entirely on localhost.

The Workflow

1. Local Polling: The backend uses imap-tools to fetch Gmail's Social category via the X-GM-RAW "category:social" search extension.
2. Sender-Aware Classification: A local regex classifier tags emails (e.g., linkedin_accepted) without needing an LLM. It strictly verifies the sender domain is linkedin.com to prevent spoofing.
3. Local Drafting: When a relevant email is found, Offmail queries the local Ollama instance running Gemma 3 1B with a strict system prompt to generate a contextual, professional reply. This takes about 10–20 seconds on a standard 4GB laptop—completely offline.
4. Routing Magic: Once the user approves the draft, the backend sends the email via SMTP, setting the To header to LinkedIn's unique [e-mail] address. This seamlessly routes the email directly into the recipient's LinkedIn DMs without violating any scraping Terms of Service.
5. Offline-First Queue: If you approve drafts while on a plane or disconnected, they sit in a local SQLite queue and automatically retry with exponential backoff once you're back online.

flowchart TB
    subgraph UserMachine["User's machine (laptop, 4GB+ RAM)"]
        UI["React UI (localhost:5173)"]
        API["FastAPI backend (localhost:8000)"]
        DB[("SQLite (inboxiq.db)")]
        Ollama["Ollama (localhost:11434)"]
        Gemma["Gemma 3 1B (open-weight model)"]

        UI <-->|HTTP /api| API
        API <-->|SQLAlchemy| DB
        API <-->|HTTP /api/generate| Ollama
        Ollama --- Gemma
    end

Why Does Open Innovation Matter?

Open innovation is the entire reason Offmail can exist securely.

If I had built this using a closed API (like OpenAI's GPT-4 or standard SaaS inbox tools like Superhuman), Arpit's private conversations with recruiters would have to sit in someone else's database. Those tools cost $20–30/month forever and stop working the moment you stop paying.

Because of open weights like Gemma 3 1B and local inference tools like Ollama, Offmail flips the script:

• Cost: $0. It runs entirely on the user's hardware.
• Privacy: Arpit's data never leaves his laptop. There is no central Offmail server.
• Customization: Want a different tone? Change one environment variable to swap models.
• Auditability: Because it's open-source, anyone can read the code to verify that their inbox isn't being harvested for telemetry.

Open source made it possible to build a powerful AI inbox assistant that respects user privacy by default.

Checkout the app here

offmail-seven.vercel.app

Prize Categories

 

✅ Best Use of Gemma

OffMail uses Gemma 3 1B running locally through Ollama to generate personalized networking follow-ups without relying on external AI APIs.
 

✅ Build for a Friend

This project was built specifically for my friend Arpit. 

 

Final Thoughts

Building OffMail made me realize something.The most exciting thing about modern AI isn't that it can write emails.It's that we can finally build useful AI tools without giving away user data.
 

No subscriptions.No telemetry.No cloud processing.
Just a laptop, an open model, and a real problem worth solving.
that's exactly the kind of future I want to build for.💙