All projects

Catchbunny

Browser testers that prove a bug before they report it. A swarm of browsers explores a running web app, replays every suspect in a fresh browser, and reports only what happens again.

Status
Shipped
Stack
Python, asyncio, Playwright, FastAPI, SQLite, MCP, React
Decisions
  • Report a bug only if it happens again Catchbunny replays every suspect in a fresh browser. What does not repeat is dropped, and what does is cut to the fewest steps that still show it.
  • Tests should outlive the model Catchbunny exports each confirmed bug as a regression test that replays with no model calls, so CI stays cheap and deterministic.
Links

How a run works

  1. Map. Find every screen and control, for each persona and device.
  2. Hunt. A swarm of browsers explores and fuzzes in parallel.
  3. Verify. Replay each candidate in a fresh browser. Only what happens again is reported.
  4. Shorten. Cut the path to the fewest steps that still show the bug.
  5. Report. Ranked, with video and logs, and as a regression test your agent or CI can run.

Map, hunt, verify, shorten, report, with the numbers from one run

Where it runs

  • Nothing goes into the app under test. No SDK and no code; it opens the app by its address, the way a person would.
  • One process serves the console, the HTTP API and the MCP endpoint on one port, so a scan a coding agent starts shows in the console as it happens.
  • After catchbunny connect, Claude Code or Codex can scan an app, read a finding, fix it and check the fix.
  • The testers need no AI. They explore and fuzz by rule. Claude or a local model through Ollama can drive them for bugs that raise no error.

Measured

On the bundled benchmark, a shop with planted bugs and traps that look like bugs, with no AI model:

Result What was measured
9 of 9 planted bugs that raise an error, found
10 of 10 confirmed findings are real, and none are reported on the fixed build
78 → 28 steps to reproduce, after shortening
10 of 10 exported regression tests fail on the buggy build and pass on the fixed one, with no model calls
4.6× faster with 8 testers than with 1

Two more planted bugs raise no error, so only a tester that understands the page can see them. With Claude Opus 5.5 driving the testers, one run confirmed both, for $0.75.

A confirmed finding: expected and actual, nine steps shortened from sixteen, and the recording of the replay with its requests