If you've used Playwright MCP for AI browser automation, you know the pain. Every page navigation dumps the full DOM tree into the model context. Simple flows like "order 5 items from this shop" can burn hundreds of thousands of tokens on navbar/sidebar/footer noise that has nothing to do with the task.
We built a small MCP layer that sits in front of Playwright and only forwards the relevant bits. Open sourced it.
📚 Full writeup: https://treesoop.com/blog/playwright-mcp-optimizer-token-saving-2026
🔧 GitHub: https://github.com/treesoop/claude-native-plugin
The problem
Playwright MCP serializes the full DOM:
AI ← {ENTIRE_DOM_JSON} ← Playwright MCP
This works for QA where you need to see everything. For "browse and take an action" it's 5-10× the tokens you actually need.
The optimizer
AI ← {relevant_only} ← Optimizer ← {full DOM} ← Playwright MCP
Three filter rules:
- Interactive elements first: button, input, a — not decorative div/span
- Semantic grouping: navigation / main / form / footer regions, so the model knows where it is
- Task-aware skipping: if the current task is "checkout", skip sidebar recommendations and ad banners
Measured impact
On a "cart → checkout" flow with GPT-4: tokens dropped substantially, and round-trip latency improved as a side effect (smaller payloads → faster agent decisions).
Not a silver bullet. For QA tasks where you need full DOM accuracy, use vanilla Playwright MCP. For general browsing / automation agents, this is the cheaper + faster path.
Tool comparison (our testing)
| Tool | Strength | Use for |
|---|---|---|
| playwright-mcp (default) | Full DOM accuracy | QA, complex validation |
| playwright-optimizer (this) | Token efficiency | Automation agents, browsing |
| vercel-browser-agent | Code generation speed | Simple browsing |
| claude-chrome-extension | Uses logged-in session | Tasks needing auth state |
We use all four for different jobs.
Install
npm install -g @treesoop/playwright-optimizer
claude mcp add playwright-opt -- playwright-optimizer
- MIT licensed
- Configurable per-site presets
-
--log-tokensflag for measurement
More OSS from TreeSoop: ai-news-mcp, hwp-mcp, whisper_transcription, claude2codex
