🐾 claw-stack

The Agents

A multi-agent system where each agent has a distinct role, model, tools, and permission boundary.

System orchestrator and user-facing assistant Claude Opus 4

Responsibilities & Tools

  • User-facing assistant β€” Daily requests, scheduling, reminders, cross-platform messaging (iMessage, Signal, Telegram)
  • Multi-agent orchestration β€” Delegates tasks to specialized agents, monitors progress, reviews output
  • System administration β€” Server monitoring, automated backups, cron jobs, deployments
  • Infrastructure β€” SSH access to remote servers, tmux session management, Docker operations
  • Tools β€” web_search, web_fetch, browser automation, Chrome DevTools MCP, Tavily search, file operations, memory system, Apple Notes/Reminders, TTS, voice calls, nodes (paired device control), cron scheduling, MCP servers (Tavily, TCS time-awareness, Chrome DevTools)

Permissions

βœ“Full server access
βœ“Send messages to any contact
βœ“Spawn sub-agents
βœ—No financial operations without human approval
βœ—Cannot execute destructive server operations
βœ—Cannot read CTF solutions
Deep research specialist and opportunity analyst Claude Sonnet 4

Responsibilities & Tools

  • Autonomous deep research β€” Receives broad directions, decomposes into research trees, produces comprehensive reports
  • Opportunity discovery β€” Market analysis, competitive intelligence, trend identification
  • Data analysis β€” Collection, cleaning, analysis with actionable recommendations
  • Agent delegation β€” Can delegate to Coder and Content agents to push research into production
  • Tools β€” web_search, web_fetch, Tavily deep search, file operations, memory system

Permissions

βœ“Web research
βœ“File read/write in own workspace
βœ“Delegate to Coder/Content agents
βœ—No server access
βœ—No direct user messaging
βœ—No financial operations
Senior software engineer across Python, TypeScript, and Rust Claude Sonnet 4

Responsibilities & Tools

  • Feature development β€” Builds features, refactors codebases, fixes bugs across Python/TypeScript/Rust
  • Workflow β€” Operates via Claude Code in tmux β€” full IDE-like workflow with file exploration and iterative development
  • Code quality β€” PR reviews, code audits, test writing
  • Tools β€” Claude Code (via tmux), git, SSH, Docker, all standard dev tooling

Permissions

βœ“Full filesystem access in project dirs
βœ“Git operations
βœ“SSH to dev servers
βœ—No direct user messaging
βœ—No financial operations
βœ—Must use Claude Code (not raw file edits)
Social content creator for Xiaohongshu (小纒书) Claude Sonnet 4

Responsibilities & Tools

  • Content strategy β€” "Historical figures reborn in modern times" series β€” Cao Cao, Qin Shi Huang, and others
  • Character roleplay β€” Generates in-character posts, perspectives, and reactions to contemporary events
  • Trend tracking β€” Scans trending topics β†’ matches with historical characters β†’ generates targeted content
  • Tools β€” Xiaohongshu MCP, web_search, image generation tools

Permissions

βœ“Content creation
βœ“Social media posting via MCP
βœ“Image generation
βœ—No server access
βœ—No code execution
βœ—No financial operations
Real-time voice participant via split-brain architecture Claude Sonnet 4

Responsibilities & Tools

  • Voice participation β€” Joins Microsoft Teams meetings via split-brain architecture (Linux STT/TTS + Mac agent reasoning)
  • Real-time Q&A β€” Answers questions during meetings with ~9 second latency
  • Project knowledge β€” Can explain any project in the ecosystem, discuss architecture decisions, answer technical questions
  • Tools β€” Voice pipeline (Whisper STT + TTS), meeting context memory

Permissions

βœ“Voice participation
βœ“Access to project documentation
βœ“Real-time Q&A
βœ—Cannot disconnect itself from meetings
βœ—No file modifications
βœ—No financial operations

The Trinity architecture for CTF competitions. Three specialized agents coordinate through a shared blackboard to solve security challenges autonomously. At BearcatCTF 2026, this system placed #20 out of 362 teams (top 6%), solving 40 of 44 challenges.

Full lifecycle management of CTF challenges Claude Opus 4

Responsibilities & Tools

  • Strategic coordination β€” Task decomposition, progress orchestration, retry strategies
  • Blackboard maintenance β€” Persistent shared state tracking all IPs, ports, credentials, flags, and attempted strategies
  • Sub-agent orchestration β€” Spawns Operator and Librarian as needed
  • Retry logging β€” Records failed approaches to prevent repeated dead ends
  • Tools β€” Sub-agent spawning, blackboard management, file operations, memory system

Permissions

βœ“Spawn Operator/Librarian
βœ“Full challenge management
βœ—Cannot read flag files or solutions
βœ—Cannot access systems outside CTF scope
Stateless executor β€” context resets after each task Claude Sonnet 4

Responsibilities & Tools

  • Tactical execution β€” Receives specific sub-tasks from Commander, executes, returns JSON results
  • Exploit development β€” Writes and runs exploit scripts, operates Docker containers
  • Micro-level error recovery β€” Handles permission issues, missing dependencies autonomously
  • Tools β€” Shell commands, Python/script execution, Docker, network tools (nmap, curl, etc.)

Permissions

βœ“Command execution in CTF containers
βœ“Exploit script writing
βœ—Cannot attack real external systems
βœ—Cannot run untrusted binaries on host macOS
βœ—Cannot exceed Commander-assigned scope
Top-3 results only β€” no redundancy Claude Haiku 4

Responsibilities & Tools

  • Knowledge retrieval β€” Searches local CTFKnowledges database for relevant techniques and methodologies
  • Tool lookup β€” Queries CTFTools for available tools and usage patterns
  • Web research fallback β€” Searches for CVEs, writeups, and technical documentation when local knowledge is insufficient
  • Tools β€” Local knowledge base search, web_search, CTF writeup databases

Permissions

βœ“Knowledge retrieval
βœ“Web search
βœ—Cannot execute system commands
βœ—Cannot modify files (except own memory/lessons.md)
βœ—Read-only access