🏗️ System Overview

┌─────────────────────────────────────────────────────────────────────────────┐ │ OPENCLAW GATEWAY │ │ (Always-on daemon process) │ │ Port 18789 (local) │ └─────────────────────────────────────────────────────────────────────────────┘ │ ┌───────────────────────┼───────────────────────┐ ▼ ▼ ▼ ┌───────────────┐ ┌───────────────┐ ┌───────────────┐ │ 💬 Webchat │ │ ✈️ Telegram │ │ 📱 Signal │ │ (Browser) │ │ (Bot) │ │ (Optional) │ └───────────────┘ └───────────────┘ └───────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────────────────────────┐ │ 🧠 BRAIN: Claude Opus 4.6 │ │ (Anthropic API - Primary Model) │ │ │ │ Planning & Reasoning ─────► Opus 4.6 │ │ Coding Tasks ─────────────► Codex 5.3 (via CLI) │ └─────────────────────────────────────────────────────────────────────────────┘ │ ┌─────────────┬─────────────┼─────────────┬─────────────┐ ▼ ▼ ▼ ▼ ▼ ┌───────────┐ ┌───────────┐ ┌───────────┐ ┌───────────┐ ┌───────────┐ │ 🔧 Tools │ │ 📁 Files │ │ 🌐 Web │ │ ⚡ Agents │ │ ⏰ Cron │ │ Layer │ │ Memory │ │ Layer │ │ Layer │ │ Scheduler │ └───────────┘ └───────────┘ └───────────┘ └───────────┘ └───────────┘

🔧 Core Components

🧠 Brain AI

  • Primary: Claude Opus 4.6 (Planning)
  • Coding: Codex 5.3 (Implementation)
  • Vision: Image analysis built-in
  • Voice: Groq Whisper (STT)

🔧 Tools EXEC

  • exec: Shell commands, scripts
  • read/write/edit: File system access
  • browser: Puppeteer automation
  • message: Channel messaging

🌐 Web SEARCH

  • web_search: Brave Search API
  • web_fetch: URL content extraction
  • browser: Full page control
  • Screenshots: Visual capture

⚡ Sub-Agents SPAWN

  • sessions_spawn: Isolated workers
  • sessions_send: Cross-session messaging
  • Parallel: Up to 8 concurrent
  • Isolated: Own context per agent

📁 Memory PERSIST

  • MEMORY.md: Long-term knowledge
  • memory/*.md: Daily session logs
  • SOUL.md: Personality/identity
  • USER.md: User preferences

⏰ Scheduling CRON

  • Heartbeats: Periodic checks (~30m)
  • Reminders: One-shot at time
  • Recurring: Cron expressions
  • Background: Overnight processing

🔄 Request Flow

1
Message Received
Gateway receives message from Webchat, Telegram, or other channel
2
Context Loaded
SOUL.md, USER.md, MEMORY.md, and recent memory files loaded
3
Brain Processing
Claude Opus 4.6 analyzes request, decides on tools/approach
4
Tool Execution
exec, browser, web_search, or other tools called as needed
5
Response Sent
Reply routed back through originating channel
6
Memory Updated
Important context written to daily log (memory/YYYY-MM-DD.md)

🎯 Model Routing

Task Type Model Provider Status
Planning & Reasoning Claude Opus 4.6 Anthropic Active
Coding & Implementation Codex 5.3 OpenAI Active
Speech-to-Text Whisper Groq Active
Text-to-Speech ElevenLabs ElevenLabs Available
Fast Inference Llama 3 Groq/Cerebras Available
Image Generation GPT-Image / DALL-E 3 OpenAI Needs Key

📁 Workspace Structure

~/.openclaw/workspace/ # Main workspace ├── MEMORY.md # Long-term curated knowledge ├── SOUL.md # Personality & identity ├── USER.md # User preferences ├── AGENTS.md # Operating instructions ├── TOOLS.md # Local tool notes ├── HEARTBEAT.md # Periodic check tasks │ ├── memory/ # Daily session logs │ └── 2026-02-09.md │ ├── brain-dumps/ # Raw ideas to process ├── overnight/ # Tasks for while you sleep │ ├── in-progress/ │ └── completed/ │ └── github-projects/ # Active code projects ├── interview-buddy/ ├── mission-control/ ├── omni-trader/ └── security-lab/

🎯 Skills (Modular Capabilities)

💻 Development

  • coding-agent: Spawn Claude Code, Codex
  • github: gh CLI wrapper
  • gemini: Google Gemini CLI

🎤 Media

  • video-frames: ffmpeg extraction
  • openai-whisper: Speech recognition
  • sag: ElevenLabs TTS

🔐 Security

  • healthcheck: System hardening
  • 1password: Secrets management

🛠️ Utilities

  • weather: Forecasts (no API)
  • skill-creator: Build new skills
  • nano-banana-pro: Diagram generation

🧠 Multi-Agent Coordination (Symlink Architecture)

Agents don't send messages to each other. They read from the same page.
One brain, many hands.

🔗 Symlink-Based Shared Context

Inspired by Eric Osiu's multi-agent architecture. All agents have /home/pato/clawd/shared-context/ symlinked into their workspace. 

shared-context/
├── priorities.md          # Current focus (from voice standups)
├── agent-outputs/         # Each agent drops work here
│   ├── orchestration/     # Pato, Ops coordination
│   ├── security/          # Sentinel findings
│   ├── content/           # Scribe drafts
│   ├── development/       # Code changes
│   └── ...
├── feedback/             # Ionu's approvals/rejections
│   └── feedback-2026-Wxx.json
├── kpis/                 # Live metrics
│   └── current.json
├── calendar/             # Today's meetings
├── content-calendar/     # Content planning
└── roundtable/          # Cross-agent synthesis

📖 Agent Workflow

  • Before every action:
  • 1. Read priorities.md
  • 2. Check agent-outputs/
  • 3. Review feedback/
  • After completing work:
  • 4. Write to agent-outputs/{domain}/

🎯 Coordination Example

  • Sentinel (7 AM): Finds security issue
  • → Writes to agent-outputs/security/
  • Ops (7:30 AM): Reads Sentinel output
  • → Creates action item
  • Atlas (8 AM): Reads Ops item
  • → Implements fix
  • Result: Zero human coordination

✅ Benefits

  • Simple: No APIs, just files
  • Observable: ls agent-outputs/ shows activity
  • Recoverable: Messages persist
  • Fast: Same inode = instant sync
  • Scalable: Add agent = ln -s + done
  • Language-agnostic: Any process can read/write

💡 Feedback Loop

When Ionu approves/rejects something:

  • Decision → feedback/feedback-{week}.json
  • ALL agents read feedback in next run
  • One decision teaches entire team

"Ionu prefers simple, file-based solutions over complex frameworks" — Pattern learned by all agents

📊 Current Status

  • Agents linked: 11/21
  • Pato 🦆, Ops 🎯, Sentinel 🔒, Oracle 📈, Watchdog 🧪
  • Blueprint 🏠, Scribe ✍️, Radar 📡, Spark 🎓, Atlas 🏗️, North 🧭
  • Shared inode: 9196302 (priorities.md)
  • Implementation: ~20 minutes
  • Lines of code: 0 (just symlinks)
  • Test status: ✅ All passed