atac 0.1.3

An Agent Trajectory Recording and Replay tool. Provides a standardized declarative Trajectory DSL alongside atac cli, atac mcp, and atac skills. Currently supports MCP, Bash, and some Agent built-in tools, recording dynamic Agent tool call trajectories into reusable static ATaC resources.

ATaC (Agentic Trajectory as Code)

English | 中文

---

中文

ATaC (Agentic Trajectory as Code) 是一个专为 AI Agent 设计的轨迹录制与回放工具。它提供了一套标准化的声明式智能体轨迹 DSL，并提供了 atac cli、atac mcp 和 atac skills。目前的轨迹 DSL 已经支持了 MCP、Bash 和部分 Agent 的内置工具，将动态的 Agent 工具调用轨迹录制为可复用的 ATaC 静态资源。

🛠 核心能力

• 声明式控制流: 在 YAML 中实现循环 (for) 和条件 (if-else) 逻辑。
• 多协议支持: 统一驱动 mcp:// (Model Context Protocol), bash:// (Shell) 及 kimi:// (Kimi 内置工具)。
• Agent 编程接口: 提供 CLI 指令，使 LLM 能够通过命令行自主构建、修改和执行复杂工作流。
• 可视化调试图谱: 内置 Web UI (atac ui) 提供工作流运行时的可视化追踪调试功能。

📋 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
Kimi / Moonshot	kimi://	✅ 已支持	支持 Kimi-CLI 所有的内置工具
Claude Code	-	🚧 规划中	待集成内置工具集

🤖 自动化构建示例 (CLI)

Agent 可以通过以下指令序列自主生成 lookup.yaml 轨迹文件：

# 1. 初始化并定义输入变量
atac init lookup.yaml --name "GeoSearch"
atac add-input lookup.yaml --name provinces --type list

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

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

# 4. 预览生成的结构
atac show lookup.yaml

🚀 快速开始

1. 安装

uv tool install atac  # 推荐
# 或 pip install atac

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.yaml 中
steps:
  - id: call_sub
    type: action
    action: bash://run
    args:
      command: atac run child.yaml --input city="Beijing"

3. 配置 MCP 环境

export ATAC_MCP_SERVER_CONFIGS="path/to/mcp_config.json"

3. 执行轨迹

atac run example/multi_province_center.yaml

4. 作为 MCP Server 启动 任何支持 MCP 的客户端 (如 Claude Desktop 或 Cursor) 均可将 ATaC 作为工具集连接：

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

如果尚未在全局安装，也可以使用 uvx 直接运行（推荐）：

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

---

English

ATaC (Agentic Trajectory as Code) is an Agent Trajectory Recording and Replay tool. It provides a standardized declarative Trajectory DSL alongside atac cli, atac mcp, and atac skills. The trajectory DSL currently supports MCP, Bash, and some Agent built-in tools, recording dynamic Agent tool call trajectories into reusable static ATaC resources.

🛠 Key Features

• Declarative Control Flow: Native for and if-else support within YAML.
• Multi-protocol Support: Unified execution for mcp://, bash://, and kimi://.
• Agentic Authoring: CLI-first design allowing LLMs to programmatically build, refine, and execute their own workflows.
• Visual Graph Debugging: Built-in Web UI (atac ui) for runtime workflow visualization and minimal-design editing.

📋 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
Claude Code	-	🚧 Roadmap	Built-in tool integration pending

🤖 Authoring Flow (CLI)

Agents can generate a lookup.yaml trajectory via direct CLI commands:

atac init lookup.yaml --name "GeoSearch"
atac add-input lookup.yaml --name provinces --type list
atac add-for lookup.yaml --in '${inputs.provinces}' --item province
atac add-action lookup.yaml --at 0 --id geo --action "mcp://amap/maps_geo" --args '{"address": "${variables.province}"}'

🚀 Quick Start

1. Installation

uv tool install atac

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.yaml
steps:
  - id: call_sub
    type: action
    action: bash://run
    args:
      command: atac run child.yaml --input city="Beijing"

3. MCP Configuration

export ATAC_MCP_SERVER_CONFIGS="path/to/mcp_config.json"

3. Run

atac run example/multi_province_center.yaml

4. Run as MCP Server Any MCP-compatible client (like Claude Desktop or Cursor) can connect to ATaC to author and run trajectories:

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

If not installed globally, you can also use uvx directly (Recommended):

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

License

MIT