A curated, transferable set of Claude Code agents, organized into small installable packages by theme. This isn't an export of everything — it's a hand-picked slice of a larger working agent fleet, kept small on purpose. Quality and coherence over volume: every agent here earned its place because it's general-purpose, stands on its own, and isn't tied to one specific deployment.
Pull the package folder you need and drop its .md files into your own project. Done.
| Package | Agents | Purpose |
|---|---|---|
orchestration/ |
13 | Building, routing, and running other agents, plus a general-purpose nine-stage spec-driven development pipeline (plan → architect → implement → integrate → test → review → security → optimize → recover). |
engineering/ |
6 | General-purpose software engineering: backend, frontend, database, TypeScript, testing, code quality. |
security/ |
6 | Application security, privacy compliance, and bug-bounty report writing. |
devops/ |
5 | CI/CD, containers, nginx, log analysis, site reliability. |
claude-code/ |
1 | A configuration helper for Claude Code itself — status line, output styles, preferences, workflow customization. |
31 agent files across 5 packages (code-quality-engineer is intentionally shipped in both engineering/ and security/ since it's directly relevant to both workflows — 30 agents, one deliberate duplicate).
Each package has its own README.md: a short "what this is for," a table of every agent with a one-sentence description of what it does and when to reach for it, and notes on anything that was considered for inclusion and left out (with the reason).
- Pick a package folder — or a single agent
.mdfile — that matches what you're working on. - Copy the file(s) into your project's
.claude/agents/directory. - That's it. Claude Code picks up any
.mdfile dropped into.claude/agents/as an available agent; no build step or registration required.
Agents are independent — you don't need the whole package to use one agent out of it. Read the agent's frontmatter and opening section before relying on it; a few reference conventions (like a task-tracking protocol) that make the most sense if your project has something similar in place, but the core capability works standalone.
Two tools that started life in this fleet outgrew "agent definition" and became their own standalone projects — linked here, not copied into this tree:
- opensource-pipeline — a fork/sanitize/package pipeline for open-sourcing a project; also merged into a 233k-star community Claude Code repository.
- ansi-stream-guard — a 130-line ANSI-escape-boundary reassembler, with 12 tests.
This is a curated release, not a full export. Some agents and entire agent families were excluded on purpose:
- Claude Code's own internal-architecture agents — agent coordination internals, MCP implementation, tool execution pipeline, prompt-assembly, and safety-architecture detail — not meant for external release, plus anything sourced from leaked material — excluded categorically.
claude-code/ships only the one configuration-focused agent that doesn't touch that internal surface. - A larger bug-bounty-hunting agent family, except the one report-writing agent (no offensive tooling) — the hunting/auth/recon/automation/orchestration members of that family are excluded.
- Several domain-specific agent families (music production, political analysis, and similar) — out of scope for a Claude Code engineering-agent release.
Package-level READMEs note anything that was specifically requested for that package and left out, with the reason (usually: superseded by a newer agent already in the package, or genuinely doesn't exist anywhere in the source collection).
These agents are extracted from a working private platform and generalized for public use — they aren't written from scratch as a demo. Expect the occasional rough edge as a result; a sanitization pass runs after this one to scrub anything deployment-specific that slipped through.
MIT — see LICENSE. Contributions welcome — see CONTRIBUTING.md.