Use AgentX from your code editor

Code editors with a built-in assistant (Claude Code, Cursor, VS Code and Windsurf) can use extra tools that another program offers. The way they talk to that program is called MCP, the Model Context Protocol, and the program is called an MCP server or a tool server.

AgentX ships one: agentx serve --stdio. Once your editor knows about it, its assistant can check your agents, read recent chats, look at schedules, and, if you allow it, send messages and hand tasks to your agents. Here is the whole path:

Choose a tool set

A tool set decides which tools the editor gets. You pick it with --tools:

Tool set	What the assistant can do
full (the default)	Everything: read, and also send messages, start tasks, change schedules, answer approvals and edit the wiki.
read	Look only: agents, health, recent chats, events, schedules, the voice queue, the wiki and project details. It sends nothing, starts nothing and changes nothing.

Start with read if you only want the assistant to see what your agents are doing. With read, a tool outside the set is refused even when the assistant asks for it by name.

Before you start

1. Terminal: run agentx --version to check that AgentX is installed.
2. Terminal: run agentx daemon status to check that your AgentX node (the background program that runs your agents) is up.
3. Note the folder AgentX is installed in, the one that holds agentx.json. The examples below use ~/agentx; use your own.

The tool server finds your node and its token by itself:

• Address. It reads node.bind from agentx.json. Without one, it uses http://localhost:18800. Set AGENTX_DAEMON_URL to reach a node on another machine.
• Token. Every call to the node carries Authorization: Bearer <token>. The token comes from AGENTX_TOKEN if you set it, then MESH_TOKEN (from the environment or the install's .env file), then dashboard.token in agentx.json. A node on the same machine accepts calls without a token; a node on another machine needs one.

Add it to Claude Code

The quickest way is the AgentX plugin. It adds this tool server and two AgentX skills with one command: Install the AgentX plugin in Claude Code. To add only the tool server:

1. Terminal: run this command, with your own install folder:

   claude mcp add agentx -- agentx serve --stdio --tools read --cwd ~/agentx

2. Terminal: to reach a node on another machine, add its address and token instead:

   claude mcp add agentx -e AGENTX_DAEMON_URL=http://node.example.com:18800 -e AGENTX_TOKEN=<your-token> -- agentx serve --stdio --tools read

3. Terminal: start a new Claude Code session.

Add it to Cursor

1. Open ~/.cursor/mcp.json in a text editor (create it if it does not exist).

2. Paste this, with your own install folder:

   {
     "mcpServers": {
       "agentx": {
         "command": "agentx",
         "args": ["serve", "--stdio", "--tools", "read", "--cwd", "~/agentx"]
       }
     }
   }

3. Save the file.

4. Restart Cursor.

Add it to VS Code

1. Open .vscode/mcp.json in your project (create it if it does not exist).

2. Paste this, with your own install folder:

   {
     "servers": {
       "agentx": {
         "type": "stdio",
         "command": "agentx",
         "args": ["serve", "--stdio", "--tools", "read", "--cwd", "~/agentx"]
       }
     }
   }

3. Save the file.

4. Restart VS Code.

Add it to Windsurf

1. Open ~/.codeium/windsurf/mcp_config.json in a text editor (create it if it does not exist).
2. Paste the same block as for Cursor (the one that starts with "mcpServers").
3. Save the file.
4. Restart Windsurf.

To reach a node on another machine from Cursor, VS Code or Windsurf, add an env entry next to args:

"env": { "AGENTX_DAEMON_URL": "http://node.example.com:18800", "AGENTX_TOKEN": "<your-token>" }

To give the assistant every tool, replace "read" with "full", or leave out --tools and its value.

Check it worked

1. Open your editor's list of tool servers (in Claude Code, type /mcp).
2. Check that agentx shows as connected.
3. Ask the assistant: "Use agentx_health to check my AgentX node."
4. The answer is your node's health report. An error instead means the tool server is running but cannot reach the node: see below.

If something is wrong

• agentx is not in the list. Check the file for a missing comma or bracket, save it, and restart the editor.
• "fetch failed". The node is not running, or it listens elsewhere. Run agentx daemon status in a terminal, or set AGENTX_DAEMON_URL to the node's address.
• "Unauthorized". The node wants a token. Set AGENTX_TOKEN to the node's MESH_TOKEN.
• "… is not in the "read" tool set". That tool acts on your agents. Change --tools read to --tools full if you want to allow it, then restart the editor.
• The editor runs an older AgentX. Run agentx --version in a terminal and update if needed; --tools arrived with this page.