131+ drop-in Claude Code subagents across 10 categories, with intelligent Opus/Sonnet/Haiku routing and granular tool permissions per role.
Before you start
Claude Code at Level 4 and at least one agent or tool-calling session under your belt. The subagent configs are markdown files, no additional installs required.
The collection is organized by role: frontend, backend, security, devops, data, and others. Each agent has a README describing its scope, what it handles well, and what it explicitly does not do. Read the 'Does not do' section. It tells you when to use a different agent.
Each agent is a single markdown file. Copy it into `.claude/agents/` in your project root. That is the only installation step. Claude Code discovers it there at session start.
In a Claude Code session, call the agent by name and give it a narrow task, one file or one feature, not the whole codebase. Subagents produce generic output when given broad inputs.
Open the agent's file and add a short section at the bottom: your stack, your conventions, and assumptions the agent should make. A security reviewer that knows you use Supabase RLS will give you more relevant findings than a generic one.
The collection's strongest move is parallel research: you name several targets, Claude spawns one subagent per target, then merges findings into a comparison. The prompt below fixes four output points per competitor so the parallel work converges into a scannable table, not five disconnected reports.
Subagent lifecycle, copy, invoke, and customize from the collection.
| Command / Skill | What it does | When to use it |
|---|---|---|
.claude/agents/ | Drop any agent markdown file here, Claude Code discovers it at session start with no other install step | Installing any agent from the collection |
@[agent-name] [task] | Invoke a specific agent for a scoped task in a Claude Code session | Narrow, domain-specific work: security review, DB schema, API design, performance audit |
Agent README | Each agent's 'Does not do' section lists explicit scope limits | Read before invoking, tells you when this agent is the wrong choice |
Project Context section | Append your stack, conventions, and baseline assumptions to the agent's markdown file | First thing after copying, a generic agent without stack context gives generic output |
Scope a task tightly | Pass one file or one feature to a subagent, not the whole codebase | Every invocation, broad inputs produce broad, unfocused findings |