I’m curious to hear if anyone here has experience using AI tools to help develop custom Frappe apps.
I’ve been reading a bit about this, and tools like Claude Code seem to get good feedback. However, the pricing is on the higher side, and I’m still unsure how far those plans will take you in a real-world codebase.
To me, this feels like a great use case for AI—maybe even for speeding up debugging. But I’d love to hear from people who’ve actually tried it:
Which AI tools have you used?
What plan did you choose?
How well did they understand Frappe/ERPNext concepts?
Did they actually improve your development speed or reduce friction?
I use Perplexity daily. I’ve created a custom space and feed with docs and urls. Then it can search through docs or code on github. Now I can ask each fields purpose or anything. It also helps a lot when coding.
I also use Kilo Code in VSCode. I’ve created some skill to help analyze and test apps.
I am fairly happy with Claude’s models but free ones also does the job.
I have used Codex from OpenAI. Lately, it has become quite good. Developped a whole education ERP using it. It takes time but it gets better and better at each iteration. It helps to have very unambiguous agents.md file and authoritative notes and documentations on what you want. So I would say yes AI is really a huge help. Careful as it doesn’t prevent security issues, it doesn’t prevent UI mess and doesn’t prevent drift (alignment between your intentions and the code). Like for everything, there is a learning curve there and the more you use it the better you’ll become at using it.
I have been playing around with it quite a lot in the past weeks. In my experience: Gemini is useless, Opus 4.6 and Codex 5.3 are quite good. GLM-5 is by far the best value for the money. Waiting for Deepseek V4 to drop.
You can get quite a lot done when developing new Apps. I have not yet tried to alter code of Apps that are already running in production - a bit scary to let AI loose on Code you have already running in production. Not sure how to handle that. I believe for AI it makes sense to build modular solutions, keep the Apps small, only let them work in one App.
I bought an Antigravity subscription (only 2 dollars for the first 3 months). There you get some Opus 4.6 credits and quite a lot of Gemini. I have a 20$ ChatGPT subscription (not enough if you code e a lot) and a z.AI subscription. Both of dem I use with opencode inside VS code.
Probably gonna try a Claude Subscription. But I heard you get way less usage than with OpenAI.
Antigravity you can use in a mode where it just runs and runs - without asking you for input - but so far the output of that has mostly been unusable. Plus antigravity is super buggy. Generally all models have quite a lot of errors. That is they crash / you don’t get responses / you have to reprompt. Not sure if it’s because of my location or my connection. In my country internet can be sketchy.
So the best usage I get out of Agents, is when I first plan a feature with the agent in opencode plan mode, make a clear task list and let the agent implement and test and then clean up after it. Since I have just been playing around with it for Prototypes and such, I haven’t looked at the code that much. But if the code will go in production, I think it’s important to check at least the diffs and make sure what happened makes sense.
Does anyone have experience with vibe coded apps running in prod? Would be interested to hear if someone vibe codes bug fixes in production.
Ah, and I either watch them work or let them run in docker, so they don’t erase my OS
