🌌 Seiza (星座)

Native TypeScript AI Orchestration Engine & MCP Server with Real-Time Bento Web Dashboard.

Seiza acts as the high-performance execution layer for Head AI Architect agents (such as Antigravity IDE and Oh My Pi), coordinating specialized autonomous sub-agents (Planner, Coder, Reviewer, Scout, Librarian, Tester, Designer, Security) using OpenAI-compatible endpoints (e.g. 9Router) with parallel DAG execution, multi-agent consensus validation, sandboxing, and interactive Human-In-The-Loop (HITL) authorization.

---

graph TD
    Parent[Antigravity IDE / Oh My Pi Agent] <-->|MCP Stdio Protocol| Server[Seiza MCP Stdio Server]
    Server <-->|Real-Time REST Task Sync| Daemon[Seiza HTTP Dashboard Daemon]
    Server <-->|Native Engine| Core[DAG Scheduler & Engine]
    Core -->|Topological Sort| DAGRunner[DAG Parallel Runner]
    DAGRunner -->|Parallel Spawns| Runner[Sub-Agent Execution Loop]
    Runner <-->|LLM / Streaming API| Router[9Router Model Router & Fallback]
    Runner <-->|Coder/Reviewer Dialogue| Consensus[Consensus Manager]
    Runner <-->|Isolated Commands| Sandbox[Sandbox Layer]
    Runner <-->|Proxy MCP Tools| Bridge[MCP Bridge Server]
    Bridge <-->|FTS5 Memory & Graph| Amneshia[Amneshia Memory Hub]
    Bridge <-->|Code Discovery| CodebaseMCP[Codebase Memory MCP]
    Daemon <-->|SSE & REST API| Dashboard[React 18 + Vite Bento Dashboard]

---

✨ Core Features

• ⚡ DAG-Based Parallel Orchestration: Automatically breaks complex coding prompts into topological dependency graphs and executes non-dependent steps in parallel.
• 🤝 Multi-Agent Peer Review (Consensus Engine): Enforces automated Coder-Reviewer dialogue loops to verify diffs and safety before applying changes.
• 🔄 Real-Time Cross-Process Task Sync: Synchronizes task execution states across standalone Stdio MCP processes and the HTTP daemon via POST /api/tasks/sync and SSE streams.
• 🛡️ Dynamic Model Router & Auto-Fallback: Integrates with 9Router daemon. Automatically retries failing models (404/429/5xx) with zero-downtime exponential backoff fallback strategies.
• 🌉 Universal Downstream MCP Bridge: Seamlessly bridges tools from Amneshia (SQLite FTS5 Long-Term Memory Hub), Codebase Memory MCP, and Context7.
• 🔒 Human-In-The-Loop (HITL): Intercepts destructive commands or tasks containing #butuh-manusia, pausing execution until approved via the Bento Web Dashboard.
• 📊 Real-Time Bento Web Dashboard: Premium React 18 + Tailwind UI featuring live DAG graphs, SSE streaming logs, agent directive editors, and token counters.
• 💾 Context Inflation Shield: Logs complete sub-agent execution trails into local SQLite (sessions.db) while returning concise abstractions to the parent agent.

---

⚡ Quick Start

1. Standard MCP Server Setup (Stdio Mode)

Add Seiza to your MCP configuration file (mcp_config.json or ~/.omp/agent/config.yml):

{
  "mcpServers": {
    "seiza": {
      "command": "wsl.exe",
      "args": [
        "-e",
        "env",
        "PATH=/home/murtix/.local/share/fnm/node-versions/v24.18.0/installation/bin:/usr/bin:/bin",
        "/home/murtix/.local/share/fnm/node-versions/v24.18.0/installation/bin/seiza"
      ]
    }
  }
}

2. Standalone HTTP Server & Web Dashboard

Run the server with the web dashboard enabled on port 3456:

seiza --http --port 3456

Or run in background daemon mode:

seiza --http --daemon

Open your browser at http://localhost:3456 to access the Seiza Web Dashboard.

---

🛠️ MCP Tools Reference

Seiza exposes the following tools to parent agents over standard MCP transports:

Tool Name	Description	Key Arguments
run_seiza_task	Executes a complex task using topological DAG scheduling & sub-agent loops.	prompt, model, cwd, dag, skills
run_single_agent	Runs a targeted autonomous sub-agent directly.	agentName, prompt, model, cwd, skills
list_seiza_agents	Lists all available agent profiles and frontmatter specs.	(none)
list_seiza_models	Fetches available 9Router models and current role assignments.	(none)
get_task_status	Retrieves real-time execution status of active or completed tasks.	taskId (optional)
list_seiza_skills	Returns all installed skills in ./skills or ~/.seiza/skills.	(none)
install_seiza_skill	Installs a skill package from GitHub or a local directory path.	source
list_bridge_tools	Discovers tools exposed by connected downstream bridge servers.	serverName (optional)

---

🔌 Downstream MCP Bridge Integration

Seiza can proxy tools from downstream MCP servers directly into sub-agent execution loops:

• Amneshia: Zero-external-DB SQLite FTS5 long-term memory hub & knowledge graph.
• Codebase Memory MCP: Graph-based structural code discovery, Cypher queries, and trace paths.
• Context7: Official up-to-date documentation engine for modern frameworks.

Manage bridge servers dynamically in the dashboard under the Bridge tab or in ~/.seiza/config.json.

---

💻 CLI Options

seiza [options]

Options:
  -V, --version        output the version number
  --data-dir <path>    custom data directory (default: "~/.seiza")
  --http               enable HTTP/SSE server mode (default: true)
  --no-dashboard       disable HTTP Web Dashboard server
  -p, --port <number>  port number (default: 3456)
  -d, --daemon         run server in background daemon mode (default: false)
  -h, --help           display help for command

---

🛠️ Development & Building

# Install root dependencies
npm install

# Typecheck and build backend
npm test
npm run build

# Build Web Dashboard
cd dashboard
npm install
npm run build

---

📄 License

MIT © Sabil Murti