Skip to main content

​
Sandbox

The Sandbox is a scratchpad for quick, throwaway agent runs. Nothing is saved to the database. No agent record is created. Use it to:
  • Test a system prompt before assigning it to a real agent
  • Try a one-off task without cluttering your agent list
  • Experiment with different models or MCP configurations

​
How it works

Configure

Set the model, system prompt, and whether to enable MCP servers. Everything lives in the left panel.

Converse

Type a prompt and get a response. The conversation appears as a chat thread in the right panel.

​
Session persistence

The Sandbox supports multi-turn conversations within a session. After the first response, muxAI captures the Claude CLI session_id and passes --resume <sessionId> on subsequent messages. The agent picks up where it left off with full context. Click New Session to reset and start fresh. Sessions are also cleared on page reload.
The Sandbox is fully stateless from the database’s perspective. No runs, no logs, no results are persisted. When you close the page, it’s gone.

​
Chat

Chat is similar to the Sandbox but designed for open-ended conversations rather than agent testing. It provides a clean chat interface for interacting with Claude directly. Both Sandbox and Chat are found in the sidebar under their respective sections.