Navital Gym
Navital Gym is a set of practice tasks for AI agents, built around everyday Indian jobs. It shows how well agents do on them. It is open for anyone to use, and live now.
git clone https://github.com/Navitallabs/navital-gym.git
cd navital-gym
pip install -e .
navital-gym run --env upi --agent oracleNavital Gym has five worlds. In each one, an AI agent gets a task and some tools, then gives an answer. Trip Planning, UPI Desk, GST Desk, Form Desk and Bhasha Support Desk (Form Desk also has a harder multi-document set). A written rule checks every answer. No AI is the judge. The code and results are public, and we add new results as runs finish.
- Live
Trip Planning
Multi-city trips under budget, rail quota, closing days and food needs.
- Live
UPI Desk
Payments, limits, failed payments and scam requests.
- Live
GST Desk
Invoices, input credit, due dates and late fees.
- Live
Form Desk
Identity forms, addresses and Indian number and date formats.
- Live
Bhasha Support Desk
Shop support in Hindi and Hinglish, plus an experimental Tamil, Bengali and Marathi split.
- How Navital Gym grades AI agents: written rules: Every Navital Gym task is graded by named, deterministic rules that recompute the answer from the sandbox. No AI judge. Here is how it works.
The code is free to use and change (Apache-2.0). The tasks are free to use too (CC BY 4.0). You can run it on your own model and send us the results. The code is here: github.com/Navitallabs/navital-gym
Leaderboard: live, updated as runs finish.
- Watch or star the repo on GitHub to be told of updates
- Email us to be told when the full report is out
- The rules are a documented simplification of public Indian practice, not live systems. Each environment page lists its simplifications and the date they were last checked.
- Names such as UPI, GST, IRCTC and Indian Railways describe the workflow. Navital Labs is not affiliated with NPCI, RBI, GSTN, UIDAI, India Post, CBIC, Indian Railways, IRCTC or any bank, and nothing here is financial, tax or travel advice. These are simplified rules for a test. They are not the law and not advice.
- Data is synthetic or from public sources. No customer data is used.
- Leaderboard rows come only from runs graded with our grader, each with the model name, version and date. Rows are added as runs finish, so the board is not complete yet.
Rules last checked: 25 October 2026
Live since 27 October 2026. New results are added as runs finish. The full report comes out on 10 November 2026.
Found a wrong rule, or want an environment built for your workflow? Contact us or write to dakshbathla@navitallabs.com.
