Assistant MCP

A separate surface: not a channel for users but access for an external agent. A client such as Claude Code connects to a workspace over the MCP protocol and works through tools — lists workflows, fetches a graph, edits it, validates it, runs it and reads the logs.

Enabling it

Settings → Assistant MCP: generate a key (shown once) and copy the ready-made client configuration. The key belongs to the workspace, not to a user; runs the agent makes are billed to the workspace owner like any other.

Permissions

Access is read-only by default: the agent sees workflows, executions and logs but changes nothing. Write permissions are granted explicitly and separately, and the server both hides the tools a key may not use and re-checks the permission at call time — knowing a tool’s name buys nothing.

What is available

Workflows (list, read at three levels of detail, create, save a version, deploy, delete, validate, run, cancel), executions (list, detail, logs, queue health), workspace data (knowledge bases, document collections, files, chat history, connections without secrets) and a dry-run template render.

What next

• The Deploy tab — channels for people rather than agents.
• Settings — where the key lives.