I audited the biggest Claude Code agent repos

The best-known Claude Code agent repositories on GitHub are free, open source and genuinely good. wshobson/agents publishes over 200 agents, plus skills and commands, under MIT; VoltAgent/awesome-claude-code-subagents collects 100+ subagents organised by domain; 0xfurai and rshah515 maintain further community collections; and hesreallyhim/awesome-claude-code is a curated list with over 50,000 stars that already shows per-entry license badges. If you want maximum selection at zero cost, clone wshobson/agents and prune. What an audit surfaces is not a quality gap — it is volume, cross-repo duplication and license ambiguity once you aggregate from a dozen sources.

I installed the large collections into a scratch project and looked at the four things that actually bite later: whether the license is clear at the source, whether the repo is maintained, how much it duplicates other repos, and what it costs your context window. Here is the side-by-side.

The repos, side by side

What each repo does well

wshobson/agents is the reference collection: broad, coherent, MIT-licensed so the redistribution terms are unambiguous, and maintained by one author which keeps the style consistent. VoltAgent's list is the best-organised by domain, so if you know you want a backend or data specialist you can find one fast. 0xfurai is useful when you want one subagent per language or framework and don't mind the breadth. And hesreallyhim/awesome-claude-code is the map of the whole ecosystem — it is where you go to discover what exists before you decide what to keep.

None of these are bad. That is the honest headline, and it matters more than any pitch: the free ecosystem is strong, and most of it is permissively licensed.

Where each one bites

The problems only show up once the files are in your .claude/ directory. Three recurring ones:

• Duplication across repos. Clone two or three of these and you get several code-reviewer agents, several security-auditor agents, several debugger agents — near-identical prompts with different names. Nothing tells you which is best, so you keep all of them.
• Volume you pay for silently. More agents does not mean better routing. Every agent and skill has a description Claude reads to decide what to delegate to; hundreds of them compete on every task and dilute the decision.
• License ambiguity once aggregated. A curated list can link to fifty repos, but it cannot guarantee each repo's LICENSE at the commit you pull from. Long-tail community collections frequently have no LICENSE file at all — which means all rights reserved, not free-to-ship.

What we'd actually keep — and why

Strip the duplicates and the specialists you will never open, and a small working set survives across almost every project:

• One code reviewer with a clear, opinionated rubric — not three.
• One test writer / runner that matches how your project actually tests.
• One debugger / error triage agent for the noisy work you want off your main thread.
• One or two stack-specific specialists for the languages you genuinely ship — and none for the twenty you don't.
• A couple of workflow commands (a diff review, a PR description) you run by hand.

This lines up with what experienced users report: they converge on a handful and delete the rest. The failure mode is not choosing the wrong agent — it is keeping all of them. If you are unsure which primitive a given job even wants, the decision table for skills, subagents, commands and MCP is the place to start, and the same subtractive logic applies to the skills worth keeping.

The honest part: you do not have to pay for any of this

To be explicit, because a paid product has every incentive to blur it: you can build a lean, excellent setup from free repos alone. wshobson/agents is MIT and free. hesreallyhim's list is free. If you have an hour to audit and prune, do exactly that and keep your money. Anyone who tells you the free options are inadequate is selling you something.

The case for an audited, de-duplicated kit

What a maintained kit sells is not secret agents — everything begins as the same open source. It sells the work: two hundred candidates reduced to the few that earn their place, duplicates resolved to the better version, every component's license verified at the exact source commit rather than trusted from a badge, and a non-overwriting installer that will never clobber a file you already have. That is the audit done for you, plus someone keeping it current as the sources drift. Whether that hour of your time is worth paying to skip is a real question, covered in whether a paid kit is worth it; the license side is covered in license hygiene for community agents.

Audit it yourself: a five-step pass

1. Pick one base repo, not five. Start from wshobson/agents so your style stays consistent and duplication starts at zero.
2. Verify the license at the source. Open the LICENSE file at the commit you are pulling from — not the README badge, not an awesome-list column.
3. Install into a scratch project first. See how many agents you actually trigger over a week of real work.
4. Delete on sight. Any agent or skill you have not used, and any duplicate of one you have, comes out. Fewer descriptions, sharper routing.
5. Record what you kept. A short CREDITS note — component, source, commit, license — turns a pile of copied files into something you can maintain and, if you work at a company, defend.

Do that and you will have a setup most paid kits would be happy to ship. The only thing you are trading is the time to run the audit — and the discipline to keep running it every time the sources change.