atac 0.3.0

Recording and Replay AI Agent Trajectory.

ATaC — Agentic Trajectory as Code

English | 中文

---

中文

ATaC (Agentic Trajectory as Code) 提供了一套专为 AI Agent 设计的声明式轨迹录制与回放接口。它作为一个轻量级的协议载体，通过统一的协议总线（Protocol Bus）无缝接入标准 MCP 服务，以及市面上各大 Agent 应用与框架的原生内置工具。通过 ATaC，开发者能将智能体运行时的动态调用流，固化为可流转、可复用、可精确重构的静态代码资产。

🛠 核心能力

• 轨迹录制: 将智能体非结构化的工具调用历史，持久化为标准化的静态 .yaml 资产。
• 精确回放: 搭载轻量化执行引擎，按序、精确还原复杂环境下的工具执行序列。
• 声明式控制流: 在 YAML 轨迹中原生支持循环 (for) 和条件 (if-else) 逻辑编排。
• 多协议总线: 统一调度 mcp:// (Model Context Protocol)、bash:// 等多源执行环境。

📋 执行器兼容性

执行器	协议	状态	说明
MCP	mcp://	✅ 已支持	原生支持所有标准 MCP 服务
Bash	bash://	✅ 已支持	支持本地终端命令及脚本执行
Claude Code	claude://	🚧 规划中	待集成内置工具集

🤖 自主构建与流转示例

1. 自动化构建 (CLI)

Agent 可以通过以下指令序列自主生成 GeoSearch 工作区轨迹：

# 1. 初始化并定义输入变量 (自动创建 .atac/GeoSearch 目录结构)
atac init GeoSearch --description "GeoSearch workflow"
atac add-input GeoSearch --name provinces --type list

# 2. 注入逻辑结构 (For 循环)
atac add-for GeoSearch --in '${inputs.provinces}' --item province

# 3. 在指定位置插入动作 (支持路径寻址)
atac add-action GeoSearch --at 0 --id geo --action "mcp://amap/maps_geo" --args '{"address": "${variables.province}"}'

# 4. 预览生成的结构
atac show GeoSearch

2. 嵌套轨迹调用 (Sub-Workflows)

ATaC 原生支持通过 bash://run 调用其它 ATaC 文件，从而实现模块化与依赖复用：

# 在 parent 轨迹的 index.yaml 中
steps:
  - id: call_sub
    type: action
    action: bash://run
    args:
      command: atac run child_workspace --input city="Beijing"

🚀 快速开始

1. 安装与环境配置

uv tool install atac  # 极简模式（推荐）

# 挂载您的任意外部 MCP 服务器环境配置：
# (支持配置多个配置文件路径，使用逗号分隔)
export ATAC_MCP_SERVER_CONFIGS="path/to/mcp_config_1.json,path/to/mcp_config_2.json"

2. 以 Skills 形式引入

将项目中的 skills/atac 目录复制到您 Agent 工作区的 skills/ 目录下即可激活：

cp -r path/to/ATaC/skills/atac ./skills/

3. 以 MCP Server 形式引入

在支持 MCP 协议的应用程序配置中添加 ATaC：

{
  "mcpServers": {
    "atac": {
      "command": "uvx",
      "args": ["atac", "mcp"],
      "env": {
        "ATAC_MCP_SERVER_CONFIGS": "path/to/mcp_config_1.json,path/to/mcp_config_2.json"
      }
    }
  }
}

---

English

ATaC (Agentic Trajectory as Code) provides a set of declarative trajectory recording and replay interfaces tailored specifically for AI Agents. Acting as a lightweight routing layer, it employs a unified protocol bus capable of seamlessly connecting to strictly standardized MCP servers alongside the proprietary built-in tools of various Agent applications and frameworks. Through ATaC, developers can persist an agent's dynamic execution flow into modular, reusable, and deterministic static code assets.

🛠 Key Features

• Trajectory Recording: Persists unstructured agent tool invocations into standardized, static .yaml assets.
• Precise Replay: Powered by a lightweight runtime engine to predictably execute complex tool sequences.
• Declarative Control Flow: Native for loop and if-else condition routing directly within the YAML schema.
• Multi-protocol Bus: Unified execution pipeline bridging mcp://, bash://, and various platform APIs.

📋 Executor Support

Executor	Scheme	Status	Note
MCP	mcp://	✅ Supported	Native support for all MCP servers
Bash	bash://	✅ Supported	Local shell commands and scripts
Kimi / Moonshot	kimi://	✅ Supported	Full Kimi-CLI toolset support (requires [kimi] extra)
Claude Code	claude://	🚧 Roadmap	Built-in tool integration pending

🤖 Authoring & Workflow Examples

1. Authoring Flow (CLI)

Agents can generate a GeoSearch trajectory workspace via direct CLI commands:

# Create a new workspace at .atac/GeoSearch
atac init GeoSearch --description "GeoSearch workflow"
atac add-input GeoSearch --name provinces --type list
atac add-for GeoSearch --in '${inputs.provinces}' --item province
atac add-action GeoSearch --at 0 --id geo --action "mcp://amap/maps_geo" --args '{"address": "${variables.province}"}'

2. Nested Trajectories (Sub-Workflows)

ATaC supports executing other ATaC files natively via the bash://run executor, allowing you to build modular, reusable sub-workflows:

# Inside parent workspace's index.yaml
steps:
  - id: call_sub
    type: action
    action: bash://run
    args:
      command: atac run child_workspace --input city="Beijing"

🚀 Quick Start

1. Installation & Environment Setup

uv tool install atac  # Minimal installation (Recommended)

# Mount your external MCP server configurations:
# (Multiple config files are supported, separate paths with a comma)
export ATAC_MCP_SERVER_CONFIGS="path/to/mcp_config_1.json,path/to/mcp_config_2.json"

2. Introduce as Skills

Copy the skills/atac directory from this repository into your Agent workspace's skills/ directory:

cp -r path/to/ATaC/skills/atac ./skills/

3. Introduce as an MCP Server

Add ATaC to the configuration of any MCP-compatible application:

{
  "mcpServers": {
    "atac": {
      "command": "uvx",
      "args": ["atac", "mcp"],
      "env": {
        "ATAC_MCP_SERVER_CONFIGS": "path/to/mcp_config_1.json,path/to/mcp_config_2.json"
      }
    }
  }
}

License

MIT