By skyforbes 
Gemini 3 dropped this week, and I've been testing it with a tool I built for browser debugging. Turns out terminal agents are really good at terminal tools. Who knew? Your agent can write code, but can't see what happens when it runs. SPAs are invisible to curl. Puppeteer means writing entire scripts just to peek at a page. bdg opens a persistent connection to Chrome via DevTools Protocol. Your agent sends atomic commands, the browser responds, and the session stays alive.
Why CLI instead of MCP?
Token efficiency: CLI tools were in the training data. MCP definitions are 5-10k before you invoke anything.
Composability: Pipe to jq, grep, whatever. MCP servers are rigid endpoints.
Transparency: Exit codes tell the story. The agent sees errors immediately and self-corrects.
vs. Puppeteer:
Not a test framework. No 50-line scripts. Just grab data and go. Session persists for iterative debugging.
Built for agents, useful for humans:
- Semantic output (70-99% token reduction vs raw HTML)
- Self-documenting CDP API (300+ methods discoverable)
- Accessibility tree inspection
- Network telemetry and HAR export
Repo: https://github.com/szymdzum/browser-debugger-cli
Works on macOS/Linux. Windows support is coming.
Would love feedback from anyone building agent workflows or tired of wrestling with Puppeteer for simple tasks.