MCP Monday: Playwright MCP
Lets your coding agent drive a real browser. Click, type, screenshot, test. Great for frontend work.
https://t.co/9PSkhqco98
Friday tip: keep your gaal.yaml in your dotfiles.
New laptop? Clone, run gaal sync, and every agent comes back with its skills and MCP servers wired.
Restore, not rebuild.
Early call we still love: gaal upserts, never overwrites.
It merges new entries into your agent config. Your hand-tuned settings stay.
Overwriting was simpler. It also eats a config exactly once.
Honest question.
Last time you added one MCP server, how many config files did you edit by hand to get it everywhere?
Reply with your number. We bet nobody's is "one."
Skill Spotlight: the Karpathy guidelines
Make your agent code like Andrej: tasteful, legible, small. Think before coding, simplicity first, surgical changes.
Every agent wants your project instructions under its own filename.
gaal maps one source to each agent's expected spot. Write the rules once, not once per agent.
Another coding agent launched recently. Looks great, I'll try it.
It's also one more config format, one more path, one more place to re-add your skills.
Better agents, worse plumbing.
Your agent skills are dependencies. Treat them like one.
gaal pins a skill, repo, or MCP to a revision. Upstream changes it? Your agents don't move till you bump the pin.
Before gaal sync touches disk, run --dry-run.
You get the full diff: every skill, every MCP, every agent it would change. Nothing written.
Read the plan, then run it.
Editing config by hand is fine. You edit package.json by hand.
The problem is editing the same MCP into four configs by hand.
One source of truth: fine. Six copies of one fact: that's the drift.