Atac

ATaC is distributed as a Python package and is best installed with the uv tool manager, which creates an isolated environment and makes the 'atac' command available globally.

uv tool install atac

In your agent application, import the ATaC service and register both your local LangChain tools and any MCP tool clients. All registered tools become available to agent-authored graph workflows.

from atac import get_service
from langchain_mcp_adapters.client import MultiServerMCPClient

service = get_service()

# Register local tools
from myapp.tools import search_tool, calculator_tool
service.register_langgraph_tools([search_tool, calculator_tool])

# Register MCP tools
mcp_client = MultiServerMCPClient({"filesystem": {"command": "npx", "args": ["@modelcontextprotocol/server-filesystem", "."]}})
service.register_langgraph_tools(mcp_client.get_tools())

Run your agent as normal through a task. ATaC intercepts tool calls and records them as a trajectory that can be replayed. The recording captures the exact sequence of tool invocations, inputs, and outputs.

Use service.run_graph to load and execute a previously recorded trajectory. Provide the module path to your graph builder function and the initial state. ATaC replays the tool calls deterministically.

from myapp.bootstrap import get_service

service = get_service()
result = service.run_graph(
    "myapp.graphs:build_graph",
    {"query": "What is the weather in São Paulo?"}
)

To use ATaC through Claude Desktop, add it to your MCP configuration so Claude can trigger graph executions conversationally.

{
  "mcpServers": {
    "atac": {
      "command": "atac",
      "args": ["serve"]
    }
  }
}