Specter — Ghost Protocol Agent Runtime

Autonomous AI agent runtime. Self-hosted. Telegram-controlled. Security-first.

What makes Specter different?

Most agent runtimes are thin wrappers around API calls. Specter is built by a cybersecurity company — security isn't a feature, it's the foundation.

| Feature | Specter | SwarmClaw | CCCBot | OpenClaw | |---------|---------|-----------|--------|----------| | Prompt injection detection | ✅ | ❌ | ❌ | ❌ | | Cost anomaly alerts | ✅ | ❌ | ❌ | ❌ | | Execution replay/audit | ✅ | ❌ | ❌ | ❌ | | Auto-lockdown on threats | ✅ | ❌ | ❌ | ❌ | | Approval gates | ✅ | ✅ | ❌ | Partial | | Programmable workflows | ✅ | ❌ | Partial | ✅ | | Health monitoring | ✅ | ✅ | ✅ | ✅ | | Telegram control | ✅ | ❌ | ✅ | ❌ |

Features

🛡️ Threat-Aware Mode

• Prompt injection detection — 14 heuristic patterns catch jailbreaks, system prompt extraction, destructive commands
• Cost anomaly alerts — per-task and daily spend limits with automatic blocking
• Rate limiting — per-minute and per-hour task creation limits
• Auto-lockdown — critical threats trigger full system lockdown until manual /unlock
• Audit trail — every action, every decision, logged and searchable

📼 Ghost Replay

• Full execution traces — every LLM call, response, fact, and decision recorded
• Markdown export — shareable, human-readable execution reports
• Time-travel debugging — replay any past task step by step

🔮 Rituals (Programmable Workflows)

• YAML-defined workflows — multi-step task sequences with dependencies
• Cron scheduling — morning-brief at 8am, security-sweep every Monday
• Step interpolation — pass results between steps with {{step_name.result}}
• Failure policies — stop, continue, or notify on step failure

💓 Haunt Mode (Daemon)

• HTTP health endpoint — GET /health and GET /metrics
• Heartbeat — periodic status updates to Telegram
• PID management — proper daemon lifecycle
• Systemd service — auto-generate service files for production deployment

📱 Telegram Control (18 commands)

Core: /run, /status, /recent, /cancel Intelligence: /memory, /cost, /audit Replay: /replay, /replays Rituals: /rituals, /ritual Security: /threats, /unlock, /approve, /deny System: /health, /ping, /start

Plus: send any message directly to queue a task.

Quick Start

git clone https://github.com/ghost-protocol-lk/specter
cd specter
npm install
npm run build

cp .env.example .env
# Add: ANTHROPIC_API_KEY, TELEGRAM_BOT_TOKEN, TELEGRAM_ALLOWED_USERS

npm start

Architecture

specter/
├── src/
│   ├── core/
│   │   ├── agent-loop.ts     # Claude-powered task execution
│   │   ├── memory.ts         # SQLite + FTS5 (tasks, facts, audit, costs)
│   │   ├── ghost-replay.ts   # Execution trace recording
│   │   ├── threat-aware.ts   # Prompt injection + anomaly detection
│   │   ├── rituals.ts        # YAML workflow engine + cron
│   │   ├── haunt.ts          # Daemon, health, heartbeat
│   │   ├── task-queue.ts     # Priority queue
│   │   ├── config.ts         # Zod-validated YAML config
│   │   └── logger.ts         # Secret-redacting logger
│   ├── channels/
│   │   └── telegram.ts       # 18-command Telegram bot
│   └── providers/
│       └── anthropic.ts      # Claude API with cost tracking
├── rituals/                   # YAML workflow definitions
├── SOUL.md                    # Agent personality
└── specter.config.yaml        # Runtime configuration

License

MIT — Ghost Protocol (ghosts.lk)