I spent a few weeks reading every Claude Code agent, skill, and command I could find. Then I built a directory and a CLI to make them installable. Here's what the data taught me.
Why I started counting
Every time I wanted to add an agent or a skill to a project, I did the same dance. Find a repo. Read the README. Copy a file. Rename it. Drop it in .claude/. Fix the frontmatter because the format was slightly off. Run it. Watch it fail. Fix it again.
The resources existed. They were just scattered across GitHub gists, blog posts, and Discord pins. None of them agreed on structure. None of them installed cleanly.
So I started a spreadsheet. The spreadsheet became a database. The database became AgentsCamp — a curated, format-validated directory of Claude Code resources you can install with npx agentscamp add. No more manual wiring.
The six categories
After indexing, the collection settled into six buckets. Here's how the 389 resources break down:
- Agents (49) — specialized subagents you hand a job to. Code review, test writing, migration.
- Skills (38) — reusable capabilities Claude can invoke on its own when the moment fits.
- Guides (99) — the biggest category. How-tos, patterns, and hard-won workflows.
- Tools (113) — actually the largest. CLIs, MCP servers, and integrations that extend what Claude can reach.
- Commands (29) — slash commands that package a prompt into one reusable invocation.
- Glossary (61) — the vocabulary. Terms you need to even read the rest.
Patterns I kept seeing
Quality varies wildly. For every agent that was tight and well-scoped, there were three that were a vague prompt in a trench coat. "You are a helpful senior engineer" is not an agent. It's a wish.
Format and validation matter more than anything. A huge share of what I found simply did not install. Wrong frontmatter keys. Missing fields. A skill that referenced a file that wasn't in the repo. This is why everything in AgentsCamp is format-validated — if it's listed, it installs and runs. That single constraint cut the collection down hard, and made what remained trustworthy.
The best resources are narrow and specific. The standout agents do one thing. A migration agent that only handles one framework version. A skill that does exactly one transform. The broad "do everything" agents were consistently the weakest. Specificity is the signal.
Duplication is everywhere. I lost count of how many near-identical "commit message writer" commands I found. People keep reinventing the same ten utilities because there was no canonical place to find them. That, more than anything, told me the directory needed to exist.
What surprised me
I expected agents to dominate. They get the attention. But tools and guides outnumbered them two to one. The real work people share isn't the flashy autonomous agent — it's the integration that connects Claude to their stack, and the writeup explaining how they got it working.
The other surprise was how much the good stuff hides. Some of the sharpest resources I found had almost no stars and no audience. They were buried in a personal dotfiles repo. Curation, not creation, turned out to be the scarce thing. I wrote up the best agents, skills, and commands once I had enough signal to rank them honestly.
I also came in fuzzy on the line between an agent and a subagent, and indexing forced me to get precise. If you're hazy on it too, I broke down subagents separately — it changed how I categorized half the agents.
Where AgentsCamp goes next
The directory is the foundation. The interesting part is what validation unlocks. Because every resource is checked, I can do things a list of links can't — surface what actually works together, flag the duplicates, and let the CLI compose a starter set in one command.
Next I want versioning, so an installed resource can tell you when its upstream has improved. And I want community submissions that run through the same validation gate, so the bar holds as the collection grows. The whole point is that you stop maintaining a personal pile of copied files and start pulling from a source that's been checked for you.
Indexing 389 resources mostly taught me that the ecosystem isn't short on ideas. It's short on curation and a format that holds. That's the gap I'm building into.
Try it
If you've been wiring up your .claude/ directory by hand, stop. Browse the directory at AgentsCamp, find what you need, and run npx agentscamp add to drop it straight into your project — format-validated, ready to run. Copy. Paste. Ship.