Claude, Cursor, and other AI assistants
Overview
Section titled “Overview”Architect lives inside the ElevenLabs dashboard, but the agent APIs behind it are also reachable from external AI assistants, such as Claude and Cursor, through two separate integrations: the hosted MCP server and the ElevenLabs plugin.
Hosted MCP server
Section titled “Hosted MCP server”The hosted MCP server is a remote Model Context Protocol server at https://api.elevenlabs.io/v1/mcp that exposes agent management tools to any MCP-compatible client. Once connected, an assistant can create, configure, and manage the agents in your workspace through natural language, with nothing to install.
It is published in the Claude Desktop connector directory, and works with any other MCP client that supports OAuth-authenticated remote servers, including Cursor.
Architect Plugin
Section titled “Architect Plugin”Architect's own skills — the same workflows it uses in-app to explore an agent, edit its configuration, workflow, and tools, write tests, and work the triage queue — come packaged in the ElevenLabs plugin, alongside the hosted MCP server and ElevenLabs's general product skills. Installing the plugin gives a coding assistant already working in your codebase the same skills Architect uses, so it can manage agents as code.
In Cursor, install the plugin from the Cursor marketplace. In Claude Code, install it with:
/plugin marketplace add elevenlabs/plugin
/plugin install elevenlabs@elevenlabsChoosing between them
Section titled “Choosing between them”Use the hosted MCP server when you want conversational agent management from a chat client with no local setup, for example Claude Desktop. Use the ElevenLabs plugin when a coding assistant is already working in your repository and should manage agents as code, with Architect's own skills guiding how it edits tools, workflows, and tests.
Approvals in these clients
Section titled “Approvals in these clients”Each client controls its own approval behavior for tool calls. Claude and Cursor, for example, let you choose which tools run automatically and which require your confirmation before every call, independently of the approval modes described for the in-app Architect chat.