Great write-up, but looking at the results, it seems the massive 78% token reduction comes primarily from the execute tool batching actions, rather than DevTools being inherently lighter than Playwright. If someone added a similar sandboxed execute tool to Playwright-MCP to allow batching, wouldn't the token usage and turn count be nearly identical? I'm trying to understand if the protocol (CDP) is the real winner here, or just the smart API design of your MCP
Yeah, you’re right. The most of the gain is from the execute tool (batching many steps in one call), not from CDP vs Playwright. If Playwright MCP added a similar sandboxed “run script that calls other tools” and the model used it the same way, token use and turn count would likely get much closer. So the win is API design (batching + scoping + lean tool defs), not the protocol. The post is about showing that this kind of design matters, we’re not claiming our stack is inherently lighter.
We ran the same verification task with Claude Code, same prompt, 3 runs per MCP. Browser DevTools MCP: ~330K tokens, 12 turns, ~2m45s. Playwright MCP: ~1.5M tokens, 48–52 turns, ~6m. So 78% fewer tokens, 76% fewer turns, ~57% faster. We compare task-level totals rather than single-turn context size, and the post explains the methodology and how we designed for fewer round-trips (execute tool, scoping). Link in the post.
Great write-up, but looking at the results, it seems the massive 78% token reduction comes primarily from the execute tool batching actions, rather than DevTools being inherently lighter than Playwright. If someone added a similar sandboxed execute tool to Playwright-MCP to allow batching, wouldn't the token usage and turn count be nearly identical? I'm trying to understand if the protocol (CDP) is the real winner here, or just the smart API design of your MCP
Yeah, you’re right. The most of the gain is from the execute tool (batching many steps in one call), not from CDP vs Playwright. If Playwright MCP added a similar sandboxed “run script that calls other tools” and the model used it the same way, token use and turn count would likely get much closer. So the win is API design (batching + scoping + lean tool defs), not the protocol. The post is about showing that this kind of design matters, we’re not claiming our stack is inherently lighter.
We ran the same verification task with Claude Code, same prompt, 3 runs per MCP. Browser DevTools MCP: ~330K tokens, 12 turns, ~2m45s. Playwright MCP: ~1.5M tokens, 48–52 turns, ~6m. So 78% fewer tokens, 76% fewer turns, ~57% faster. We compare task-level totals rather than single-turn context size, and the post explains the methodology and how we designed for fewer round-trips (execute tool, scoping). Link in the post.
[dead]