star-orchestrator
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
star-orchestrator
一个本地优先、可自控的智能体编排框架:把「大模型决策」与「真实动作执行」彻底分层, 让 LLM 只负责选技能与填参数,所有副作用都经过显式技能注册表 + 风险确认闸门后才发生。
编排层用 Python 标准库
http.server起服务(零框架依赖);执行层用 FastAPI/uvicorn,只接受
skill_id + args,不做意图解析;MCP / A2A 协议接入全部手写 JSON-RPC over Streamable HTTP,不引入官方 SDK;
内置 SSRF/越权/危险命令拦截、长期记忆、本地 LLM 故障兜底、SSE 流式与心跳。
用户 / 外部智能体 / Bot
│
▼
┌──────────────────────────┐ ┌──────────────────────────┐
│ orchestrator :8700 │ │ exec :8730 │
│ 唯一对话层 │ │ 技能引擎 │
│ · Function Calling 循环 │──────▶ │ · config 注册表驱动 │
│ · SSE 流式 + 心跳 │ skill │ · risk=confirm 二次确认 │
│ · 参数护栏 arg_guard │ + args │ · 危险命令/敏感路径拦截 │
│ · 长期记忆 / 本地 LLM 兜底│ │ · ResultEnvelope 统一返回 │
└──────────────────────────┘ └──────────────────────────┘
│ mcp_ 前缀 ▲
├─────────────▶ 外部 MCP Server │
│ a2a_ 前缀 │
└─────────────▶ 外部 A2A Agent │
config/super_connector.json目录结构
star-orchestrator/
├── orchestrator/ 编排层(8700)
│ ├── main.py 入口
│ ├── config.yaml 非敏感配置(凭据一律走环境变量)
│ ├── api/gateway_api.py HTTP 网关 + FC 工具循环 + SSE
│ ├── adapters/ LLM / 执行层 / 可选外部大脑适配器
│ ├── router/ 路由、参数护栏、tools schema、MCP、A2A
│ ├── storage/ 会话记忆 + 长期记忆(SQLite)
│ ├── services/ 平台化(审计/计价/健康)+ 通用查询服务
│ └── utils/ 配置加载、trace
├── exec/ 执行层(8730)
│ ├── server.py FastAPI 入口:鉴权 / 技能 / SSE
│ ├── auth.py 多用户认证(PBKDF2 + 会话 + 模块授权)
│ └── super_connector.py 技能引擎(handler 注册表 + 风险闸门)
├── config/super_connector.json 技能注册表(唯一扩展点)
├── examples/custom_skill.py 最小可运行的自定义技能示例
├── desktop/ 桌面壳源码(Swift + WKWebView / Tauri)
├── scripts/start.sh 一键启动两层
├── scripts/oss/audit_public.py 敏感信息闸门(CI 与发布前共用)
└── tests/ 176 个用例(全离线可跑)Related MCP server: CONTINUUM
快速开始
git clone https://github.com/<your-account>/star-orchestrator.git
cd star-orchestrator
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env # 按需填入 DEEPSEEK_API_KEY 等凭据
./scripts/start.sh # 同时起 8700(编排层)与 8730(执行层)验证:
curl http://127.0.0.1:8730/api/health # 执行层(loopback 免登录)
curl http://127.0.0.1:8730/api/skills # 技能清单(由注册表派生)
curl http://127.0.0.1:8700/health # 编排层双引擎状态跑测试(无需任何外部服务与网络):
python3 -m pytest tests/ -q # 176 passed
python3 scripts/oss/audit_public.py # 敏感信息闸门端口表
端口 | 服务 | 说明 |
8700 |
| 唯一对话层:路由、FC 循环、SSE、记忆 |
8730 |
| 执行层:技能清单 / 执行 / 确认闸门 / 唤醒 SSE |
所有地址与端口均可用环境变量覆盖,详见 .env.example。
内置技能
技能由 config/super_connector.json 单一权威源定义,risk=read 免确认直接执行,
risk=confirm 必须先经二次确认闸门。
分类 | 技能 |
信息获取 |
|
出行/消费 |
|
办公 |
|
创作 |
|
桌面操作 |
|
系统 |
|
新增技能见 docs/SKILLS.md,可直接照抄 examples/custom_skill.py。
协议接入
MCP(作为 Host/客户端):在
orchestrator/config.yaml的mcp.servers白名单里声明外部 MCP Server,工具会以mcp_<server>_<tool>前缀并入 Function Calling 面板。详见 docs/MCP.md。A2A(Agent 联邦):在
a2a.servers声明外部 Agent,其技能以a2a_<agent>_<skill>前缀暴露; 默认只读委托,allow_write=false时写操作会被拒绝。详见 docs/A2A.md。
两者均为手写 JSON-RPC over Streamable HTTP,无第三方 SDK 依赖。
安全模型
分层纪律:执行层不接受自然语言,只接受
skill_id + args;意图解析归编排层。风险闸门:
risk=confirm技能生成一次性令牌,用户确认后才真正执行。参数护栏:
orchestrator/router/arg_guard.py在执行前拦截 SSRF(内网 URL)、 家目录越权读写、非法枚举取值、空参数等。命令与路径:
exec/super_connector.py内置危险命令正则与敏感目录黑名单。凭据:任何 api_key / 令牌 / 口令都只从环境变量读取,
config.yaml全为空串占位。发布闸门:
scripts/oss/audit_public.py在 CI 与发布前扫描全仓,命中敏感信息即非零退出。
⚠️ 本框架默认提供终端执行 / 文件读写 / 代码执行等高权限技能。请仅在受控环境运行, 并遵循 SECURITY.md 中的部署建议。
文档
docs/ARCHITECTURE.md — 分层、数据流、扩展点
docs/SKILLS.md — 技能开发指南
docs/MCP.md — 作为 MCP Host 接入外部 Server
docs/A2A.md — Agent 联邦与委托语义
docs/DEPLOYMENT.md — 部署与进程编排
CONTRIBUTING.md — 贡献指南
许可
MIT,见 LICENSE。
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
MCP server for your apps' tools and custom tools, plus hosted AI agents and approval-gated workflows
Zero-secret MCP gateway for AI agents: risk-scored, audited calls with human-in-the-loop approval.
Hosted MCP server and Agent Skills for specs, engineering threads, and project memory.
Governed autonomous-agent platform: HITL approvals, budget gates, audit chain, effect ledger.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceEnables offline AI agent automation with embedded local LLM (Qwen 2.5), sandboxed file operations through AgentFS, and dynamic skill loading. Exposes capabilities via MCP with tri-state safety guards for private, air-gapped environments without network connectivity or API costs.-
- AlicenseAqualityBmaintenanceUniversal verifiable recovery for long-running AI agents with semantic checkpoints, idempotent action ledger and hash chained log as a deny by default MCP server. Framework agnostic with adapters for LangGraph, LangChain and OpenAI plus gateway and OTel.1327Apache 2.0
- FlicenseAqualityCmaintenanceAn MCP server that exposes governance, trust-scoring, compliance, guardrail, cost, drift, and supply-chain scanning tools and resources to any MCP client over stdio, Streamable HTTP, or legacy HTTP+SSE. It lets agents route every tool call through a deterministic five-way decision (allow, redact, require approval, deny, or quarantine) with hash-chained evidence, human approval workflows, and in-agent trust gates for LangChain, LangGraph, and Google ADK.30-
- FlicenseNot gradedqualityBmaintenanceEnables AI agents to execute real-world API actions with exactly-once reliability, deduplication, auditing, and human-in-the-loop safeguards via MCP.3-