Agents and MCP

Agents combine instructions, a model or route, editable persistent memory, and an explicit set of tools. The workspace also tracks tasks and execution logs.

Create an agent

1. Open Agents and create a profile.
2. Write focused instructions and select a model or persistent route.
3. Add only the memory the agent should retain.
4. Connect an MCP server and discover its tools.
5. Choose a tool allowlist and approval policy before execution.

MCP server configuration

Use the server configuration appropriate to the transport supported by the application. Treat command arguments and environment variables as sensitive: never put provider keys inside memory or agent instructions.

Approval policies

Tool allowlists define what an agent may request. Approval policies determine when a human confirmation is required. Sensitive actions should always require confirmation, especially actions that write files, send messages, spend money, publish content, or alter external systems.

Execution logs and tasks

Execution logs retain the agent, tool, status and relevant result metadata for review. The task board is the persistent foundation for planning and tracking agent work; it is not an unattended cloud scheduler.