Jaanch 🔍

An AI tester that uses your web app like a real person, finds what breaks, tells you how to fix it, and writes the tests for you.

Jaanch (जाँच, Hindi for inspection) opens your site in Brave with Playwright, clicks buttons, fills forms and moves between pages. BharatCode AI decides what to try next and explains every bug it finds.

npx github:Ajay702/jaanch http://localhost:3000

_{Real run on the included ChaiCart demo with BharatCode qwen-3.8-27b. Waiting time for the model is cut out; everything else is real speed ×1.8.}

Jaanch v0.1.0  — testing http://localhost:4173
  brain: BharatCode AI (qwen-3.8-27b) · up to 30 actions

◆ /menu “Menu · ChaiCart”
   3/30 AI click button "Add Filter Coffee"
       ✖ TypeError: Cannot read properties of undefined (reading 'toFixed')
◆ /signup “Sign up · ChaiCart”
  11/30 AI click button "Create account"
       ✖ TypeError: Cannot set properties of null (setting 'textContent')
  12/30 AI click link "Dashboard"
       ✖ GET /api/orders returned 500

Found 4 high, 2 medium, 2 low across 5 pages in 13.5s
  📄 Report: jaanch-report/report.html
  🧪 Tests:  jaanch-report/jaanch.spec.js

Why

Writing end-to-end tests is the chore every team skips. So bugs like a typo'd element ID, a button wired to a function that doesn't exist, or an API that 500s only show up when a real user hits them.

Code reviewers (human or AI) read your code. Jaanch uses your app, so it catches the bugs that only appear at runtime.

What you get

🧭 Explores on its own	Follows links, clicks every button and fills forms with realistic Indian test data (names, +91 numbers, PIN codes).
🐞 Catches real bugs	JavaScript crashes, console errors, failed API calls (4xx/5xx), broken pages and images, unclickable buttons.
♿ Accessibility checks	Images without alt text, inputs without labels, buttons without names.
🧠 Explains the fix	BharatCode AI reads the error and the source line it came from, then writes the cause and a concrete fix.
🧪 Writes Playwright tests	One test per bug. It fails while the bug exists and passes once you fix it, so the bug can't come back.
📄 Shareable report	report.html with screenshots, repro steps and the full journey, plus report.json for tooling.
🛡️ Safe by default	Never clicks Delete, Log out, Pay now, Unsubscribe and the like, and never leaves your domain.

Quick start

Requirements: Node 20.12+ and Brave (Chrome or Edge also work).

# 1. Try it on the built-in demo shop with planted bugs
git clone https://github.com/Ajay702/jaanch && cd jaanch && npm install
npm run demo -- --watch          # watch Brave click around by itself

# 2. Point it at your own app
npx github:Ajay702/jaanch http://localhost:3000

Turn on the AI brain

Without a key Jaanch still works, using built-in rules. With a BharatCode key it explores smarter and writes much better explanations:

npm run set-key        # paste your key from https://bharatcode.ai/account (hidden input, saved to .env)
npx github:Ajay702/jaanch --check-ai            # ✔ BharatCode AI works (qwen-3.8-27b): नमस्ते

Students get free access (Qwen, 10M tokens a day). A full 25-step run on the demo shop uses about 20 model calls and roughly 12k tokens.

Options

jaanch <url> [options]

  -s, --steps <n>        Max actions to try (default 25)
  -w, --watch            Show the browser while it works, with a live HUD
  -r, --record <file>    Save a video of the run (.webm), HUD included
  -o, --out <dir>        Report folder (default ./jaanch-report)
  -t, --tests <file>     Where to write Playwright tests
  -b, --browser <name>   brave | chrome | edge (default brave)
      --no-ai            Rules only, no model calls
      --fail-on <level>  Exit 1 on bugs of this severity or worse (for CI)
      --open             Open the report when done
      --check-ai         Test your BharatCode key

Use it in CI

- run: npm start &                      # start your app
- run: npx github:Ajay702/jaanch http://localhost:3000 --fail-on high
  env:
    BHARATCODE_API_KEY: ${{ secrets.BHARATCODE_API_KEY }}
- uses: actions/upload-artifact@v4
  if: always()
  with: { name: jaanch-report, path: jaanch-report }

How it works

flowchart LR
    A[Open URL in Brave] --> B[Read page:<br/>buttons, links, fields]
    B --> C{BharatCode AI<br/>picks next action}
    C --> D[Playwright clicks / types]
    D --> E[Listen: JS errors,<br/>console, network]
    E --> B
    E --> F[BharatCode AI explains<br/>each bug + source line]
    F --> G[report.html +<br/>Playwright tests]

Loading

1. See. Every visible link, button and field is listed and tagged.
2. Decide. Jaanch builds a safe shortlist (fill empty fields, then submit, then untried buttons, then new pages), with dangerous actions already removed. BharatCode's qwen-3.8-27b picks one and writes realistic input. Without AI, the first item on the shortlist wins.
3. Act and listen. Playwright performs the action while Jaanch records page errors, console errors and failed requests, each tied to the exact step that caused it.
4. Explain. For each bug, Jaanch fetches the source lines around the error and asks the model for a title, severity, what the user sees, the root cause and a fix, using structured JSON output.
5. Lock it in. Each bug becomes a Playwright test that replays the steps on that page and asserts the bug is gone.

How BharatCode is used

• Endpoint: https://bharatcode.ai/api/model/v1/responses (OpenAI-compatible Responses API)
• Model: qwen-3.8-27b (override with JAANCH_MODEL)
• Structured output: text.format: json_schema, so every answer is validated JSON, never scraped prose
• Resilient: retries 503 "model busy" with backoff, detects cut-off output, falls back to rules if the model is unreachable, and never blocks a run

Project layout

bin/jaanch.js      CLI entry
src/browser.js     finds Brave/Chrome/Edge, launches via Playwright
src/explorer.js    the see → decide → act loop
src/planner.js     safe action shortlist + BharatCode planner
src/collector.js   turns browser events into deduped issues
src/analyze.js     severity + AI explanations with source context
src/testgen.js     Playwright regression test generator
src/report.js      HTML + JSON report
demo/              ChaiCart, a demo shop with 6 planted bugs

Development

npm test          # unit tests + an end-to-end run against the demo shop
npm run demo      # run Jaanch on ChaiCart

Roadmap

• Log in first using a saved Playwright storageState
• Mobile viewport and slow-3G runs
• Comment the report on GitHub pull requests
• Visual regression (screenshot diffs between runs)

License

MIT · Built for the BharatCode Build League 2026 by Ajay Raghav.