3fLeadFlow Agent

An AI agent inside Slack that captures leads in plain English, scores their urgency, drafts the reply and reminds your team until no lead is forgotten.

Comment

Inspiration

Small appointment-based businesses salons, clinics, physios, tradespeople lose customers every day for one simple reason: somebody meant to reply and forgot. I run an automation agency in Limerick, Ireland, and I kept seeing the same story. A customer messages at 9 AM, nobody answers, and by lunchtime she's booked with a competitor. I wanted an agent that makes this impossible.

What it does

3fLeadFlow Agent is an AI lead management agent that lives inside Slack. You talk to it in plain English: type "add a lead: Emma Ryan, 087 222 3344, wants a massage on Friday" and it captures the lead, scores how urgent it is, and drafts a warm reply your staff can send straight away. Ask "what leads are still new?" and it answers from its records. The part I am most proud of: a background engine watches every lead, and if one sits uncontacted too long, the agent posts a reminder in the channel by itself with waiting times and keeps reminding until someone acts. No lead ever gets forgotten.

There are quick slash commands too: /newlead, /leads, /updatelead, /assignlead.

How we built it

The whole system runs on an MCP (Model Context Protocol) server. The lead database is never touched directly by the bot it is exposed through four MCP tools: add_lead, get_open_leads, update_lead_status and assign_lead. When you mention the agent, your message goes to Qwen along with the tool definitions, and the model decides which tool to call. Nothing is hardcoded — remove the MCP server and the agent literally has nothing to work with.

Stack: Python, Slack Bolt (Socket Mode), FastMCP, Qwen (qwen-plus) on Alibaba Cloud Model Studio, SQLite, APScheduler. Deployed 24/7 on Alibaba Cloud ECS as a systemd service.

Challenges I ran into

I started this project having never built on Slack and never used MCP. My favourite bug: after deploying to the cloud, the agent silently lost all its tools it turned out the MCP server was being spawned with a Python command that exists on Windows but not on Linux. One line (sys.executable) fixed it, but finding it taught me more about agent architecture than any tutorial.

Accomplishments that I am proud of

I built this solo, in days, around a full-time job starting from zero Slack and zero MCP knowledge. The MCP integration is structural, not decorative: remove the server and the agent can do nothing. I debugged a silent cross-platform failure on a live cloud server. And it solves a real problem I've watched hurt real businesses in my community this is the working heart of a future product, not a tech demo.

What I learned

MCP clicked for me during this build. I now understand why it is becoming the standard for agent-tool architecture. I also learned that an agent is only as good as its deployment: "works on my laptop" means nothing until it works at 3 AM on a server with nobody watching.

What's next for 3fLeadFlow Agent

Block Kit interactive buttons on the reminders, a daily summary digest, and connecting it to real client enquiry channels. This agent is the working heart of a product I plan to bring to small businesses in Ireland.

Built With

Share this project:

Updates