How to adopt AI plugin for Redmine Helpdesk

#ai #llm #rag #agents

Hi, I'm Marat from RedmineUP Cloud. We're about to ship AI Agent, a Redmine plugin with a built-in MCP server. Before releasing it, we let our own support team use it on real customer tickets.

Turns out the hard part isn't the LLM. It's the order you introduce things in.

The data is already there

If your support runs on Redmine, you're sitting on years of solved tickets, wiki pages and customer history. Today that knowledge just sits there. RAG makes it usable.

Here's what we learned, step by step.

1. Answers from your own history (no automation yet)

The agent searches wiki pages and issues with Redmine's native search, as the user who asked. No separate index to sync, no permission leaks between customers. Every answer links to its source: "see #4521 and the wiki page on SSO".

2. Triage + drafts, but a human hits Send

A webhook fires when an email becomes a ticket. The agent picks a category, priority and assignee only from IDs that exist in Redmine, writes a draft, and saves everything as a private note:

PUT /issues/4812.json
{
  "issue": {
    "category_id": 7,
    "priority_id": 4,
    "assigned_to_id": 23,
    "notes": "AI triage: billing, EN, customer frustrated. Similar: #3310.\n\nDraft reply: ...",
    "private_notes": true
  }
}

Three rules that saved us:

Closed lists. Reject any value that isn't in your list.
No auto-send. Private notes only, so nothing reaches the customer.
No loops. The agent's own update fires a webhook too, so filter out bot events.

Run this for 2–4 weeks and log the share of drafts sent with no or minor edits, per category. That number decides where automation is safe.

3. Reports you just ask for

Through MCP, a support lead types: "Median and P90 first response time for project X, share answered within 4h, trend by year." The agent builds the query, the charts, and then explains the long tail.

4. A first-line agent (still in design)

Auto-replies only in narrow categories with proven accuracy: password resets, "where's my invoice". Otherwise it hands off to a human with a summary.

5. Customer 360

CRM + Helpdesk data in one call. A small but useful trick: compute scores like LTV in code, and let the LLM only explain them. The numbers stay reproducible.

The takeaway

Don't start with a chatbot. Start with retrieval, add drafts, measure, and automate only where the numbers say it's safe.

Full write-up with API details: 5 Steps to Transform Your Redmine Helpdesk with an AI Plugin

What's the first thing you'd hand to an AI agent in your support queue? 👇

DEV Community