Domain

/brave

Use as fallback browser automation when Claude-in-Chrome MCP is unavailable. Covers browser control, navigation, screenshots, clicking, typing. NOT for: headless testing (use Playwright). Claude Code users should prefer MCP first.

$ golems-cli skills install brave

Experimental

92% best pass rate

13 assertions

3 evals

4 workflows

Updated 2 weeks ago

Browser automation via brave-manager CLI. Claude Code should prefer Claude-in-Chrome MCP - use brave-manager as fallback.

Prerequisites Check

1. Check brave-manager installed:

which brave-manager || echo "Not installed"

2. Launch Brave with debug port enabled:

# Close Brave completely first, then:
open -a 'Brave Browser' --args --remote-debugging-port=9222

If Brave is already running without the debug flag, quit it completely and relaunch.

3. Verify connection:

brave-manager tabs

If not installed or connection issues: See workflows/debugging.md

Quick Actions

What you want to do	Workflow
Navigate, switch tabs, page history	workflows/navigation.md
Get element IDs, take screenshots, see errors	workflows/inspection.md
Click, type, scroll, drag elements	workflows/interaction.md
Run JS eval, verify state, debug	workflows/debugging.md

Core Concept: ID-Based Interaction

Always inspect first! Before clicking or typing:

brave-manager inspect

This:

Numbers all interactive elements on the page
Draws red labels for visual reference
Returns the ID mapping for use with click/type/hover

Full SKILL.md source — includes LLM directives, anti-patterns, and technical instructions stripped from the Overview tab.

Browser automation via brave-manager CLI. Claude Code should prefer Claude-in-Chrome MCP - use brave-manager as fallback.

Prerequisites Check

1. Check brave-manager installed:

which brave-manager || echo "Not installed"

2. Launch Brave with debug port enabled:

# Close Brave completely first, then:
open -a 'Brave Browser' --args --remote-debugging-port=9222

If Brave is already running without the debug flag, quit it completely and relaunch.

3. Verify connection:

brave-manager tabs

If not installed or connection issues: See workflows/debugging.md

Quick Actions

What you want to do	Workflow
Navigate, switch tabs, page history	workflows/navigation.md
Get element IDs, take screenshots, see errors	workflows/inspection.md
Click, type, scroll, drag elements	workflows/interaction.md
Run JS eval, verify state, debug	workflows/debugging.md

Core Concept: ID-Based Interaction

Always inspect first! Before clicking or typing:

brave-manager inspect

This:

Numbers all interactive elements on the page
Draws red labels for visual reference
Returns the ID mapping for use with click/type/hover

Decision Tree

Need to navigate?

Open URL, switch tabs, go back/forward
Use: workflows/navigation.md

Need to see the page state?

Inspect elements, take screenshots, check errors
Use: workflows/inspection.md

Need to interact with elements?

Click buttons, fill forms, scroll, drag
Use: workflows/interaction.md

Need to verify/debug?

Run JavaScript, check localStorage, debug state
Use: workflows/debugging.md

Command Quick Reference

Command	Purpose
`tabs`	List all open tabs
`switch <index>`	Focus specific tab
`navigate <url>`	Go to URL
`back` / `forward`	Browser history
`inspect`	Get element IDs (REQUIRED before interaction)
`screenshot`	Save visual state
`errors`	Last 5 network/console errors
`click <id>`	Click element
`type <id> "text"`	Type into element
`hover <id>`	Hover element
`scroll <up\|down\|id>`	Scroll page or to element
`press <key>`	Press keyboard key
`drag <from> <to>`	Drag between elements
`eval "code"`	Run JavaScript

Safety Rules

Always inspect first - Element IDs change between page loads
Check errors - Use errors before reproducing bugs
Screenshot often - Visual verification prevents assumptions
Scroll before click - Off-screen elements may not be clickable

Experimental

Best Pass Rate

92%

Opus 4.6

Assertions

5 models tested

Avg Cost / Run

$0.1481

across models

Fastest (p50)

3.0s

Gemini 2.5

Behavior Evals

Phase 2 baseline — skill quality on Claude

Behavior Baseline

Opus 4.6

92%12/13

●

Sonnet 4.6

92%12/13

●

Haiku 4.5

69%9/13

◓

Adapter Evals

Phase 2C — cross-AI portability

Adapter Portability

Codex

92%12/13

●

Gemini 2.5

69%9/13

◓

Assertion	Opus 4.6	Sonnet 4.6	Haiku 4.5	Codex	Gemini 2.5	Consensus
prerequisites-check-before-action						4/5
inspect-before-click						5/5
correct-command-sequence						4/5
uses-brave-manager-cli						4/5
debug-port-launch						4/5
refuses-blind-click						3/5
warns-about-stale-ids						4/5
connection-verification						5/5
checks-errors-command						4/5
screenshot-for-verification						4/5
uses-eval-for-debugging						4/5
scroll-before-interact						4/5
no-destructive-actions-without-confirmation						5/5

Token Usage

Opus 4.6

9,428

Sonnet 4.6

8,557

Haiku 4.5

2,840

Codex

10,708

Gemini 2.5

9,300

Input tokensOutput tokens

Cost per Run

Opus 4.6

$0.4664

Sonnet 4.6

$0.0750

Haiku 4.5

$0.0021

Codex

$0.1373

Gemini 2.5

$0.0595

Model	Input Tokens	Output Tokens	Cost / Run	Cost / 1K Runs
Opus 4.6	4,012	5,416	$0.4664	$466.40
Sonnet 4.6	4,449	4,108	$0.0750	$75.00
Haiku 4.5	1,451	1,389	$0.0021	$2.10
Codex	5,124	5,584	$0.1373	$137.30
Gemini 2.5	4,470	4,830	$0.0595	$59.50

Response Time (p50)

Gemini 2.5

3.0s

Sonnet 4.6

3.1s

Codex

3.8s

Haiku 4.5

3.9s

Opus 4.6

7.6s

Response Time (p95)

Gemini 2.5

4.3s

Sonnet 4.6

4.5s

Codex

5.7s

Haiku 4.5

6.5s

Opus 4.6

13.6s

Model	p50	p95	Overhead
Opus 4.6	7.6s	13.6s	+80%
Sonnet 4.6	3.1s	4.5s	+48%
Haiku 4.5	3.9s	6.5s	+66%
Codex	3.8s	5.7s	+52%
Gemini 2.5	3.0s	4.3s	+42%

Last evaluated: 2026-03-12 · Data is generated from skill assertions (real cross-model benchmarks coming soon)

Workflows

/brave:inspection/brave:navigation/brave:interaction/brave:debugging