Agent Integration
| Feature |
Playwright CLI |
Playwright MCP |
Stagehand |
browser-use |
Vibium |
Selenium 4 |
| CLI skill support |
Native (install --skills) |
No |
No |
No |
Native |
No |
| MCP server |
Companion (MCP Registry) |
Yes |
Via Browserbase |
No |
Built into binary (V1) |
No |
| First-party test agents |
Yes (planner/generator/healer) |
Shared |
No |
No |
No |
No |
| Agent-native design |
Yes |
Adapter |
Yes |
Yes |
Yes |
No |
| Zero-config setup |
Yes |
Mostly |
Yes |
Yes |
Yes |
No |
| Auto browser download |
Yes |
Yes |
Yes |
Yes |
Yes |
No |
Token Efficiency
Microsoft's own benchmark (the only vendor-published, like-for-like number as of July 2026):
| Metric |
Playwright CLI |
Playwright MCP |
| Typical task |
~27,000 tokens |
~114,000 tokens |
| Ratio |
~4x cheaper |
baseline |
| Long sessions |
up to ~10x reported |
— |
| Mechanism |
Snapshots/screenshots to disk, read on demand |
Full state streamed into context per action |
Stagehand and browser-use sit between the two poles: cheaper than raw MCP streaming, costlier than disk-first CLI, with exact numbers depending heavily on model and page complexity — treat any per-step token figure you see in a blog post as an estimate, not a spec.
Browser Control
| Feature |
Playwright CLI |
Playwright MCP |
Stagehand |
browser-use |
Vibium |
Selenium 4 |
| Auto-wait / actionability |
Yes (engine) |
Yes (engine) |
Yes |
Partial |
Yes (server-side in Go) |
No (manual) |
| Element addressing |
Snapshot refs (e8) |
A11y tree refs |
Natural language intent |
NL + vision |
CSS selectors + find-all |
Selectors |
| Screenshot |
To disk |
Inline |
Yes |
Yes |
Yes |
Yes |
| Network interception |
Via engine/trace (HAR in tracing) |
Yes |
Yes |
Limited |
BiDi network module |
Partial (BiDi) |
| Session/state reuse |
Named sessions, state-save/state-load |
Server-side session |
Yes |
Yes |
Daemon mode |
Manual |
| Cross-browser |
Chromium/FF/WebKit |
Chromium/FF/WebKit |
Chromium-first |
Chromium-first |
Chrome (V1) |
All major |
Page Understanding
| Approach |
Playwright CLI |
Playwright MCP |
Stagehand |
browser-use |
| Page state |
YAML a11y snapshot on disk |
Rich a11y tree in context |
Intent resolution over DOM |
Vision + DOM |
| Agent reads |
On demand (file) |
Always (streamed) |
Abstracted away |
Streamed |
| Semantic understanding |
High (a11y summary) |
High |
High |
High (vision) |
| Cost of understanding |
Paid only when needed |
Paid every action |
Per intent call |
Per vision call |
Ecosystem (July 2026)
| Aspect |
Playwright |
Stagehand |
browser-use |
Vibium |
Selenium |
| Age |
Mature (2020) |
New (2024-25) |
New (2024) |
New (2025, V1 Jun 2026) |
Veteran (2004) |
| GitHub stars |
70K+ |
~22.8K |
~81K |
Small |
30K+ |
| Production adoption |
Widespread |
Growing fast |
Large OSS community |
Not evident yet |
Universal |
| Commercial support |
Microsoft |
Browserbase |
No |
No |
Multiple vendors |
| Trajectory since Jan 2026 |
Shipped CLI + agents; recommends CLI over MCP |
Standout of the agentic-SDK wave |
Most-starred OSS agent framework |
Shipped V1; adoption unproven |
Steady 4.4x releases; v5 still pending |