ATaC provides a unified tool runtime for AI agent workflows, enabling agents to orchestrate and reuse assets through graph-based code. It targets developers building complex agent systems, offering a structured approach to manage tools and define reproducible agent execution paths, thereby accelerating development and enhancing maintainability.

How It Works

ATaC centralizes tool management by allowing registration of both local tools (defined via langchain_core.tools.tool) and remote MCP (Multi-Server MCP) tools (via langchain_mcp_adapters.client.MultiServerMCPClient) into a single AtacService instance. Agents can then define their workflows as executable graphs using langgraph, directly calling these registered tools. The runtime facilitates the execution of these compiled graph workflows, promoting the gradual development of reusable and maintainable agent logic.

Quick Start & Requirements

• Primary install / run command: uv tool install atac or pip install atac.
• Non-default prerequisites and dependencies: Python. The examples utilize langchain_core and langgraph. MCP tool integration requires setting up compatible MCP servers.
• Links:
  • PyPI: https://pypi.org/project/atac/
  • License: https://opensource.org/licenses/MIT
  • CI: https://github.com/ATaC-team/ATaC/actions

Highlighted Details

• Unified registration for local and MCP-based agent tools.
• Supports agent-authored graph workflows using langgraph.
• Enables one-command execution of defined graph workflows.
• Facilitates the creation of reusable and maintainable agent task logic.

Maintenance & Community

Active Continuous Integration is indicated by the CI status badge. No specific details regarding core contributors, sponsorships, or community channels (like Discord/Slack) are provided in the README.

Licensing & Compatibility

The project is licensed under the MIT License, which is permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

The provided documentation does not explicitly detail limitations. Integration with MCP tools requires a separate server setup. The example code suggests a specific project structure (myapp.bootstrap, myapp.graphs) which may represent a recommended pattern.