Announcing the launch of the changAI - The AI Agent for ERPNext

1

Please download and install the app from GitHub - ERPGulf/changAI: AI for Frappe ERPNext · GitHub
Detailed instruction from here https://youtu.be/fFxyAY_sVNs

ChangAI by ERPGulf: Open-Source AI for ERPNext

ChangAI is an open-source assistant built for the Frappe/ERPNext ecosystem. It empowers you to query your business data using plain, natural language. No complex report builders or SQL knowledge required.

Configuration :

After installation navigate to ChangAI Settings in your Desk to configure your engines:

1. Google Gemini (Core Query Engine)

2. Amazon Polly (Optional Voice Assistant)

If you want to “hear” your insights, go to the Details tab:

3. Schema Update (Not Required for Test)
ChangAI comes pre-configured with the Standard ERPNext Schema, meaning you can start querying core modules immediately after installation without any additional mapping.

While the core model is trained on ERPNext, you can easily sync your specific ERPNext customizations . Simply provide an Anthropic Claude API Key ( https://console.anthropic.com/) in the Remote tab, then click “Update Schema” in the Training tab to enrich the AI’s understanding of your specific environment.

Critical First Steps

To ensure the AI understands your specific data, perform these two actions immediately after installation:

  1. Download Embedding Model (Local AI): Click this button in Settings. This downloads our latest model to your local server, ensuring data privacy.
  2. Update Master Data (Mandatory): This syncs your Master Tables. Without this sync, the AI won’t recognize specific names or records in your queries.

Resources

4 Likes
AI/ChatGPT Strategy for ERPNext
2

We have tested it on FrappeCloud and working well. We have applied for Marketplace as well. Waiting for approval.

Will update here once its done.

3

Cool, Game changer…

4

Now you can speak and listen to your ERP.
Added Amazon polly and TTS options fro Speech-to-text and text-to-speech .
Please pull new update from Github.

5

Screenshot 2026-03-30 at 4.07.12 PM
Screenshot 2026-03-30 at 4.07.12 PM760×1262 120 KB

Screenshot 2026-03-30 at 4.06.58 PM
Screenshot 2026-03-30 at 4.06.58 PM746×1252 122 KB

6

It seems like no one wants to type these days! The most common feedback we received was about talking to ERPNext rather than texting it.
We heard you loud and clear.

We’ve just integrated Amazon Polly into changAI to bring you near-perfect voice synthesis. With a generous free tier of 5 million characters, it’s a total game-changer for accessibility and speed.

Prefer to keep it simple? We’ve also included standard browser-based TTL options to give you full flexibility.

Check out this short video where our colleague walks you through the setup process . watch it here

3 Likes
7

Hello Guys, Congratulations on the launch of ChangAI — it’s a very promising and impressive addition to the Frappe/ERPNext ecosystem. I have started testing it in my development environment using the Gemini free tier, and the initial experience has been quite interesting.

While exploring the system, I have encountered a few challenges, especially in a multi-company ERP setup:

  • When retrieving customer outstanding amounts, the system sometimes includes transactions even from cancelled documents, which affects accuracy.

  • The AI requires very specific customer naming; for example, I need to enter the exact full name like “ABC Consulting FZE” instead of just “ABC Consulting” for correct results.

  • In some cases, queries such as yearly profit calculations are not returning fully accurate outputs.

I understand this is still evolving, and I’m hopeful these areas will improve in upcoming updates.

Additionally, I would appreciate your guidance on improving AI accuracy and training. Specifically, I’d like to understand how to make the system more intelligent in recognizing entities despite variations in wording, and how best to utilize the training/schema update process effectively for custom ERPNext environments.

Looking forward to your suggestions and continued improvements. Thank you for your efforts on this excellent tool.

1 Like
8

@Support-at-ERPgulf

Voice addition is nice, also no one typing this days, just give some text to AI and ask to format and reply.
Also for this AI can you explain how routing works, I am getting issue sometimes, I used intent routing in
Brain — voice/text input first classifies intent, then routes to the
right tool/context/permission scope. Also sharing what I am building this days, your input will be valuable.

AI-formatted version (same point, expanded):

Sharing something I’m building in the same space — looking for Frappe
beta testers.

The problem:
Frappe stores code and config in the database (Server Scripts, Custom
Fields, Property Setters, Workflows). Claude Code, Codex, Cursor go
blind here — source of truth isn’t in the filesystem.

What I’m building:

Brain — persistent memory + live context layer. Tracks runtime
state: doctype schemas, customizations, server scripts, workflows.
Also handles intent routing — input (voice or text) gets classified
first, then routed to the right tool with the right permission scope.

Echo — AI agent on top of Brain. Multi-user, security-first,
row-level isolation, audit trail on every action.

Stack: Node.js · Next.js 15 · SQLite + vector hybrid · Cerebras +
Claude routing · MCP tools · Playwright · JWT auth · AWS EC2

Status: 40-45% end-to-end automation on custom ERPNext builds.
Playwright integration pushed UI automation to ~100%.

Looking for best ideas from Frappe/ERPNext devs on custom builds to beta test.

here is the prototype which developed using same architecture http://echo.sbknext.com

— Sambhaji