Implements comprehensive solution for browser_click and other interactive tools
returning massive responses (37K+ tokens) due to full page snapshots.
Features implemented:
1. **Snapshot size limits** (--max-snapshot-tokens, default 10k)
- Automatically truncates large snapshots with helpful messages
- Preserves essential info (URL, title, errors) when truncating
- Shows exact token counts and configuration suggestions
2. **Optional snapshots** (--no-snapshots)
- Disables automatic snapshots after interactive operations
- browser_snapshot tool always works for explicit snapshots
- Maintains backward compatibility (snapshots enabled by default)
3. **Differential snapshots** (--differential-snapshots)
- Shows only changes since last snapshot instead of full page
- Tracks URL, title, DOM structure, and console activity
- Significantly reduces token usage for incremental operations
4. **Enhanced tool descriptions**
- All interactive tools now document snapshot behavior
- Clear guidance on when snapshots are included/excluded
- Helpful suggestions for users experiencing token limits
Configuration options:
- CLI: --no-snapshots, --max-snapshot-tokens N, --differential-snapshots
- ENV: PLAYWRIGHT_MCP_INCLUDE_SNAPSHOTS, PLAYWRIGHT_MCP_MAX_SNAPSHOT_TOKENS, etc.
- Config file: includeSnapshots, maxSnapshotTokens, differentialSnapshots
Fixes token overflow errors while providing users full control over
snapshot behavior and response sizes.
Co-Authored-By: Claude <noreply@anthropic.com>
- Add browser_install_extension, browser_list_extensions, browser_uninstall_extension tools
- Support session-based extension isolation between MCP clients
- Extensions loaded via --load-extension Chrome flags at browser startup
- Browser auto-restarts when extensions are added/removed
- Validation ensures extensions only work with Chromium browser
- Warning system for Chrome channel vs pure Chromium compatibility
- Extension management persists across page navigations within session
- Updated README with complete extension tool documentation
🤖 Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
Add comprehensive artifact storage system with session-specific directories:
- Add --artifact-dir CLI option and PLAYWRIGHT_MCP_ARTIFACT_DIR env var
- Create ArtifactManager class for session-specific artifact organization
- Implement ArtifactManagerRegistry for multi-session support
- Add tool call logging with JSON persistence in tool-calls.json
- Update screenshot, video, and PDF tools to use centralized storage
- Add browser_configure_artifacts tool for per-session control
- Support dynamic enable/disable without server restart
- Maintain backward compatibility when artifact storage not configured
Directory structure: {artifactDir}/{sessionId}/[artifacts, videos/, tool-calls.json]
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Added browser_list_devices tool to show 143+ available device profiles organized by category (iPhone, iPad, Pixel, Galaxy, Desktop, Other)
- Enhanced browser_configure tool with device emulation using Playwright's device descriptors database
- Added support for geolocation coordinates with accuracy settings
- Implemented locale and timezone configuration for internationalization testing
- Added colorScheme preference (light/dark/no-preference) for accessibility testing
- Included permissions management for various browser APIs (geolocation, notifications, camera, microphone)
- Device emulation properly overrides individual viewport/userAgent settings when specified
- All context options are properly applied and browser context is recreated with new settings
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Fix browser_configure tool to properly use browser context factory
- Remove bypass of browser context factory that was ignoring configuration changes
- Headed mode now works correctly when DISPLAY is available
- Browser windows properly appear when switching from headless to headed mode
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Add browser_configure tool to change headless/headed mode, viewport, and user agent during session
- Fix STDIO entry point by preventing stdin close handlers in STDIO mode
- Fix headed mode default behavior when DISPLAY is available on Linux
- Add dynamic browser configuration update mechanism in Context class
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add three new MCP tools for recording browser automation:
- browser_start_recording: Enable video recording with configurable size/filename
- browser_stop_recording: Stop recording and return video file paths
- browser_recording_status: Check recording status and configuration
Key features:
- Automatic video saving when pages/contexts close
- Configurable video size and output directory
- Enhanced LLM guidance with clear workflow instructions
- Integration with existing browser context management
Videos are saved as .webm files in {outputDir}/videos/ directory.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add three new MCP tools for recording browser automation:
- browser_start_recording: Enable video recording with configurable size/filename
- browser_stop_recording: Stop recording and return video file paths
- browser_recording_status: Check recording status and configuration
Key features:
- Automatic video saving when pages/contexts close
- Configurable video size and output directory
- Enhanced LLM guidance with clear workflow instructions
- Integration with existing browser context management
Videos are saved as .webm files in {outputDir}/videos/ directory.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
This fixes an issue that there were sticky launch errors. When the
[following code
path](a15f0f301b/src/context.ts (L307-L339))
was throwing, the Error was stored in the Promise and not cleared
afterwards, this meant:
- If a browser was not there and the user tried to install it via
`browser_install` it was never working since the error was sticky.
- If other errors like CDP is not available yet etc. error appear a
re-connect would not work - the MCP server would require a restart.
Test plan: Since we don't have any `browser_install` tests I added a CDP
test for now to cover this bug.
Calling a tool that resolves modal state, when there's no such modal
state visible, currently shows this misleading message:
```md
Tool "browser_file_upload" does not handle the modal state.
### Modal state
```
Instead, we should show the error message from the tool implementation.
