Pin the bug, watch it live
Hunter · Testing and debugging
Matt Pocock TDD
Anthropic Webapp Testing
Chrome DevTools MCP
Drives live Chrome for clicks, console, network, screenshots, and performance traces.
How it works
Foe: Bug Slime, a defect in your code.
Foes
- Bug SlimeHaze Poison Ward Insight Flicker Pinned Enraged
Party · Context
- IllusionistYour agent100/100Haze Poison Ward Insight Flicker Pinned Enraged
The Bug Slime blocks the way.
What really happensIt stands for a defect in your code.
Illusionist: Context 100/100.
The Bug Slime uses Regress.
What really happensA fix in one place breaks something that used to work.
Illusionist takes 8. Illusionist: Context 92/100.
The Chrome DevTools orb opens a portal into live Chrome.
What really happensClaude Code or Codex starts chrome-devtools-mcp. Chrome launches when a tool needs it, not on connect.
Bug Slime takes 29.
The orb reads the console, the network, and a picture of the page.
What really happensThe agent inspects console messages with source-mapped stacks, network requests, and screenshots.
Bug Slime takes 13.
The Bug Slime uses Regress.
Illusionist takes 8. Illusionist: Context 84/100.
Puppeteer clicks through the page and waits for each result.
What really happensAutomation uses Puppeteer and waits for action results. A performance trace can be recorded in Chrome DevTools.
Bug Slime takes 32.
The Bug Slime's trail is written down from the live page.
What really happensYou get what the live Chrome session showed: console, network, screenshots, or a performance trace.
Bug Slime takes 26. Bug Slime falls.
The Bug Slime falls.
1 / 8
Learn this spell
Claude Code
claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latestCodex
codex mcp add chrome-devtools -- npx chrome-devtools-mcp@latestNotes
Needs Node.js LTS, current stable Chrome, and npm. Official support is Google Chrome and Chrome for Testing. The browser starts when a tool needs it, not when the server connects. Page contents are sent to the MCP client, so do not open accounts you would not share with the agent. Usage stats are on by default; pass --no-usage-statistics to opt out. Performance tools may call the CrUX API unless you pass --no-performance-crux.
Records
Recorded use16root sessions
Test recoveries7 of 7sessions with a failing test
Curator's machine: macOS 27.0.1 · Apple Silicon (arm64), read when these stats were published. Session logs do not record the OS.
Sessions per month
View data
| Month | Sessions |
|---|---|
| 2026-07 | 8 |
| 2026-08 | 4 |
| 2026-09 | 4 |
Harness
- Codex16 (100%)Version 0.144.0-alpha.4 in the first session, 0.155.0-alpha.9.2 in the latest
View data
| Harness | Sessions | Version |
|---|---|---|
| Codex | 16 | Version 0.144.0-alpha.4 in the first session, 0.155.0-alpha.9.2 in the latest |
Often used with
View data
| Item | Sessions |
|---|---|
| Matt Pocock TDD | 6 |
| shadcn Skill | 1 |
| Supabase Postgres Best Practices | 1 |
What a recovery means
A failed test was followed by a passing test of the same runner family in one session. It does not prove this spell caused the fix.
First recorded: . Most recent: .