Skip to main content
Back to Writing

Expansion & Compression

•4 min read
AI DevelopmentClaude CodeBuilding

el capitan hook

Every guide about extending Claude Code tells you which MCP servers to install. Almost none tell you which hooks to write. After four months of daily use, my ratio is twelve servers to one hook, and that ratio is backwards for what the tools are actually good at.

MCPs and hooks solve opposite problems. An MCP expands what Claude can do. A hook constrains how it does it. One adds capability, the other removes variance. Most people stack servers like browser extensions and never touch hooks at all.

inhale

An MCP server is a process that exposes tools over a protocol. Claude connects, reads the tool schemas, and can do things it ordinarily cannot. File systems, databases, design systems, analytics, video editing software, whatever. If the thing you use has an API, there's probably a server for it. And if there isn't, you can probably build one.

My current stack: PostHog, Supabase, Vercel, Sentry, Stripe, Figma, Canva, Gmail, Google Drive, Notion. Last night we added Adobe After Effects and Premiere Pro (which I didn't expect to exist, but it does, and it's actually pretty useful.) That's 12 servers that let Claude touch everything from the database to a video editing timeline.

The discovery problem is real, but here are three that are worth noting:

  • PulseMCP indexes thousands of servers with filters and search.
  • awesome-mcp-servers repos on GitHub (wong2, appcypher, tolkonepiu) get updated by humans who use these things.
  • GitHub search for "mcp-server" plus your tool name finds the one-off servers that never made a list.

The Premiere server showed up by searching 'premiere pro mcp' on GitHub. One developer, a few hundred stars, and it worked on the first try. There's probably a server for whatever obscure tool you use. You just have to look.

exhale

A hook is a shell command that fires on a Claude Code event. Before a tool runs, after it runs, when a session starts, when one ends. You configure them in settings.json with a matcher pattern and a command. Super easy and will save you so much hassle.

Every hooks guide online opens with the same example: auto-format with Prettier after every edit. It's the obvious one. It's also probably unnecessary if your editor formats on save, which it already does. Part of why every guide lands there is that hooks are genuinely hard to teach. The useful ones are situational, born from a bug you hit on your machine, in your project, at your stack level. It doesn't generalize well.

Here's my entire hook config, global, the only one I've written in four months:

"PostToolUse": [{
  "matcher": "Write|Edit",
  "hooks": [{
    "type": "command",
    "command": "f=$(jq -r '.tool_response.filePath // .tool_input.file_path'); [ -n \"$f\" ] && rm -f \"$(dirname \"$f\")/._$(basename \"$f\")\" 2>/dev/null; true"
  }]
}]

It deletes macOS resource-fork sidecar files after every write. Translation: this portfolio lives on an external SSD. When macOS writes to that drive, it leaves hidden ._filename companion files scattered around. Next.js tried to compile those as real files and the build fails. After annoyingly having to clean it up by hand so many times, I decided to create a hook to accommodate this bug.

That hook has run thousands of times since. I haven't thought about resource forks again. That's what a good hook does.

jenga

Adding an MCP is shopping. You browse, you install, you get a small dopamine hit from new capability. Writing a hook is janitorial. You notice a repeating annoyance, diagnose it, then write shell you'll never look at again. One feels like progress. The other feels like chores.

But shopping doesn't change your workflow. The tenth server you add runs once a week if that. The one hook you write runs every time Claude touches a file. The leverage is inverted from the effort.

The test for whether a hook is worth writing: have I fixed the same thing more than twice? If yes, it belongs in settings.json. The bug you're tired of is the hook you should have written yesterday.

press start

If you use Claude Code daily and have zero hooks, write one this week. Pick the most recent small annoyance you cleaned up by hand. That's it. You don't need the twelve-event taxonomy or a production CI pattern. You need one PostToolUse matcher and a command that runs in under a second.

If you have ten hooks and two MCPs, you're doing it backwards in the other direction. The asymmetry cuts both ways. Expansion without compression leaves a messy agent. Compression without expansion leaves a tidy agent that can't do much.