Claude Code Subagents Collection
A collection of 33 specialized AI subagents for Claude Code, covering the full-stack development lifecycle with intelligent auto-delegation and multi-agent orchestration.
Evidence shows: README suggests copying agent files to ~/.claude/agents/ and configuring multiple MCP servers (including magic requiring an API key), which require manual user action, with no indication of automatic execution or privilege escalation. No malicious code or credential theft detected. Deductions: no explicit least-privilege scope, no user confirmation mechanism, data flow not transparent, sensitive data handling not mentioned, dependency security not assessed, external effects not clarified, rollback not provided, source attribution only based on README links.
Evidence shows: README provides a structured agent format and consistent naming conventions, agents organized by category, showing good internal consistency. Deductions: dependencies on MCP servers (e.g., magic, context7) require external services, availability not verified; no failure messages or error handling guidance.
Evidence shows: README clearly defines target audience (Claude Code users) and multiple usage scenarios (automatic, explicit, hybrid), and provides capability boundaries (e.g., 'When to Use Which Agent' table). Deductions: trigger precision relies on description text, no precise trigger conditions; environment fit only for Claude Code, no other environments.
Evidence shows: README provides clear information architecture (agents organized by category), installation notes (manual and quick), naming conventions (kebab-case), examples and FAQ (usage examples and troubleshooting). Deductions: known limitations only mention token overhead, not comprehensive; versioning only implied via git, no explicit version or changelog; maintenance responsibility not clear.
Evidence shows: README provides rich usage examples and output format descriptions, agents designed to improve development efficiency, marginal value in providing specialized subagents. Deductions: cost-benefit not quantified, only token overhead mentioned, no actual performance data.
Evidence shows: README claims (e.g., '33 agents') consistent with file list, but no verification method provided. Deductions: claims lack traceability, no cross-source corroboration, facts and inferences not clearly separated.
- Configuring MCP servers (e.g., magic) requires an API key; ensure key security and do not commit to public repositories.
- Installation instructions suggest copying agent files to global directories, which may conflict with other tools; proceed with caution.
- Automatic invocation relies on description text, which may lead to unexpected selections; consider explicit agent specification for critical tasks.
What does this agent do, and when should you use it?
This repository provides a collection of 33 specialized AI subagents for Claude Code, organized into categories like development, infrastructure, quality testing, data & AI, security, and business. Each subagent is a Markdown file with standardized frontmatter (name, description, tools) and a system prompt defining its role, expertise, and capabilities. They are designed to be automatically invoked by Claude Code based on context analysis, or explicitly called via natural language. The collection includes an agent-organizer meta-agent that analyzes project requirements, assembles optimal agent teams, and orchestrates multi-phase workflows with quality gates. The README provides installation instructions, usage examples, and performance metrics from real-world scenarios. Subagents can optionally leverage MCP servers for enhanced reasoning and context, but they are not strictly required. The project is MIT licensed and scaffolded from wshobson/agents.
The repository provides a set of instruction files (Markdown) that define specialized subagents for Claude Code. When a user submits a request, Claude Code analyzes the context (keywords, file types, task type) and automatically selects the appropriate subagent for execution. Each subagent's frontmatter includes a description that triggers its invocation. The subagent's system prompt directs the AI to behave as an expert in a specific domain, such as backend architecture or code review. Users can also explicitly call a subagent by name in natural language, e.g., "Use the code-reviewer to check my recent changes". For complex projects, the agent-organizer can be enabled via CLAUDE.md in the project root to plan and coordinate multiple subagents, using MCP servers like sequential-thinking and context7 for enhanced capabilities. The output is the generated code, analysis, or documentation that the subagents produce, integrated into Claude Code's workflow.
- Full-stack developers implementing a new feature who want automated coordination between frontend, backend, and testing agents, e.g., requesting 'Implement user authentication' to trigger backend-architect, security-auditor, and test-automator.
- Code reviewers needing specialized review of pull requests, using code-reviewer and security-auditor for quality and security checks.
- Backend architects designing microservices and REST APIs, leveraging backend-architect for schema design and architect-reviewer for consistency.
- Performance engineers troubleshooting bottlenecks, using performance-engineer for analysis and database-optimizer for query tuning.
- DevOps teams responding to production incidents, using devops-incident-responder for log analysis and incident-responder for crisis management.
- Data scientists performing analysis and building ML pipelines, using data-scientist for SQL queries and ai-engineer for LLM applications.
What are this agent's strengths and limitations?
- Comprehensive coverage of development tasks with 33 specialized subagents, including frontend, backend, security, testing, and data/AI domains.
- Intelligent orchestration: Claude Code auto-delegates, and agent-organizer enables complex multi-agent workflows with quality gates.
- Each subagent has clear triggers in its description, making it easy to invoke the right expertise.
- Real-world examples with token counts and timelines provide transparency about costs and performance.
- Tightly coupled to Claude Code; cannot be used with other AI tools or APIs without modifications.
- Requires familiarity with Claude Code to manage multi-agent coordination effectively.
- Optional MCP servers add configuration overhead; without them, some capabilities are limited.
- Multi-agent workflows can incur high token usage (e.g., 850K tokens in the complex example), leading to significant costs.
- Some claims in the README (e.g., test coverage, token counts) may be exaggerated or unverified.
How do you install or deploy this agent?
- Ensure Claude Code is installed.
- Clone the repository into the agents directory:
cd ~/.claude
git clone https://github.com/lst97/claude-code-sub-agents.git(or if the directory exists, pull the latest: git pull origin main)
- Alternatively, manually copy individual agents:
mkdir -p ~/.claude/agents/lst97
cp /path/to/agents/*.md ~/.claude/agents/lst97- Verify installation:
ls ~/.claude/agents/lst97/*.mdIn Claude Code, run "List all available subagents" to confirm.
- For full performance, configure MCP servers in ~/.claude.json as documented, e.g., sequential-thinking, context7.
How do you use this agent?
Once installed, subagents are automatically available in Claude Code. Use explicit natural language requests to call specific agents, e.g., "Use code-reviewer to analyze this pull request", or rely on automatic delegation. For complex projects, copy the CLAUDE.md file to your project root to enable agent-organizer dispatch. Then issue a high-level command like "Implement user authentication" to trigger multi-agent coordination. Consult the README examples for detailed usage patterns and resource expectations.
How does this agent compare with similar options?
The repository does not explicitly compare itself to other agent collections, but it is scaffolded from wshobson/agents, which is a similar collection of Claude Code subagents.