wtf is w-bonkers?

w-bonkers turns Claude Code or Codex into a disciplined assistant for managing a stock portfolio on the Indian NSE market. The idea behind it is that most investment plans do not fail because of bad stock picks, they fail because of drift: forgetting what you decided last week, redrawing your stop losses in your head, or losing track of your own rules. This project tries to fix that by keeping your entire plan as data in a single file called state.json, rather than as a memory the AI has to reconstruct every time you talk to it. A one time installer interviews you about your goals, how much money you are investing, your risk appetite, and your existing holdings, sometimes even reading a salary PDF for tax context, and writes all of that into your plan file. After that, a refresh command you run on a schedule reads any feedback you left, pulls current market prices, applies a fixed set of rules, and only changes the plan when a rule actually triggers. Most days, nothing changes. Your buy and sell orders show up as tasks in Todoist, each one listing the buy zone, stop loss, and target price. You place the actual trades yourself in your broker's app, then check off the task and can leave a comment like "bought at 332", which the next run reads before doing anything else. The agent itself never places trades, since the market data connection it uses is read only for trading. Everything the system fetches, including prices, news, and past decisions, is saved locally, so the whole plan and its history survive even without the AI or an internet connection. To use it, you need Claude Code or the Codex CLI, a Groww connection for live prices and holdings, and a Todoist connection for the feedback loop, plus Python with a few packages for rendering and fallback prices. Setup is a git clone followed by telling your AI agent to read and run an installer prompt, which handles authorization, the interview, and scheduling, in about fifteen minutes. This is explicitly framed as an education project, not investment advice, and the author notes it is very new, having only been running on real money since July 2026. It is released under the MIT license.