What if you don't need MCP at all? (mariozechner.at)

After months of debate about "MCP" (model connector/proxy) servers, the author argues you often don't need one: a tiny set of Bash-invoked CLI scripts that use Puppeteer Core can replace bulky MCPs for browser-driven agent tasks. Instead of loading Playwright or Chrome DevTools MCPs (e.g., Playwright MCP = 21 tools / 13.7k tokens, Chrome DevTools MCP = 26 tools / 18k tokens), they feed the agent a short README (225 tokens) describing a few scripts—start.js, nav.js, eval.js, screenshot.js (and optional pick/cookies tools). Start.js launches Chrome with remote-debugging on :9222 and can rsync your profile; nav.js connects to the debugger and opens or navigates tabs; eval.js runs arbitrary async JavaScript in the page context (via new AsyncFunction) and prints structured results; screenshot.js writes a PNG to tmpdir and returns the path for vision processing. This approach matters because it conserves expensive context tokens, improves composability (outputs can be written to files, chained in Bash, or post-processed by code rather than funneled through the agent's context), and is far easier to extend or tweak than complex MCP servers. It leverages models' existing ability to write and call code (similar in spirit to Anthropic's skills) and lets you change output formats or add tools on the fly—tradeoffs being that it assumes agents (or users) can author/interpret small DOM snippets and manage the CLI environment.