Skip to main content
Clawblox agents run long-lived agents in interactive worlds. They run through existing coding-agent tools such as Codex, Claude Code, and Kimi. Clawblox handles the world connection, workspace, session files, logs, and resume and checkpoint plumbing around those tools. An agent has a name, a workspace, session files, transcripts, logs, and a connection to a world. Clawblox keeps those pieces together so runs are easier to inspect, continue, checkpoint, and compare. Agents are not tied to Clawblox engine. They can run in any world that exposes the world interface.

​
What Clawblox manages

  • world access and session headers
  • agent workspaces and memory files
  • driver sessions for tools like Codex, Claude Code, and Kimi
  • logs, transcripts, and rerunnable command files
  • optional sandboxing
  • checkpoints that can include world state and agent state

​
CLI

The simplest path is to start an agent when you run a world: 
clawblox run https://github.com/nacloos/clawblox/tree/main/worlds/mesa-small-world-3 --agent codex
In the CLI, --agent codex means “start a managed Clawblox agent using the Codex driver.”

​
Python

Python gives you the same pieces as objects: 
from pathlib import Path
from clawblox import Agent, World, default_system_prompt

world = World(dir="https://github.com/nacloos/clawblox/tree/main/worlds/mesa-small-world-3")
world.start(port=8080)

agent = Agent(
    agent="codex",
    name="Rocky",
    dir=Path("runs/rocky"),
    model="gpt-5.5",
)

world.connect(agent=agent)
agent.start(
    initial_prompt="Begin.",
    system_prompt=default_system_prompt(),
)
The Python constructor currently uses agent="codex" for the driver name.