MCP

Configure Model Context Protocol servers, refresh their tools, and share servers with your workspace.

Model Context Protocol (MCP) servers extend the agent with additional tools. Define local or remote servers in mcp.json, then use Settings to inspect their status and control workspace sharing.

Where to find it

Open Settings → MCP.

Server definitions live at:

text
~/.endos/mcp.json

The file supports both local stdio servers and remote MCP endpoints.

Controls

ControlWhat it does
AddOpens or creates mcp.json in Writer so you can edit server definitions.
RefreshReloads the server list and connection status.
Your serversShows each server's name, type (stdio or remote), and connection status.
Shared with everyoneShares a server you own with the workspace.
Shared with youShows a read-only list of servers shared by other members.

Configure mcp.json

The file must be valid JSON with a top-level mcpServers object. Each key is a server name that you choose; that name becomes part of the generated mcp_* tool names.

json
{  "mcpServers": {    "local-server": {      "command": "npx",      "args": ["-y", "some-mcp-server-package"],      "env": {        "API_KEY": "..."      },      "disabled": false    },    "remote-server": {      "url": "https://example.com/mcp",      "headers": {        "Authorization": "Bearer ..."      },      "disabled": false    }  }}

Field reference

FieldServer typeMeaning
commandstdioExecutable to spawn. Required for a local server.
argsstdioOptional list of command arguments.
envstdioOptional environment variables for the server process.
cwdstdioOptional working directory.
urlremoteMCP endpoint URL. Required for a remote server.
headersremoteOptional HTTP headers, such as authorization headers.
disabledbothSet to true to keep the definition without connecting to the server.

Each server must define either command for a stdio server or url for a remote server. An entry with neither field is invalid.

The format is compatible with MCP server entries used by Cursor and Claude Desktop.

Load tools after editing

After saving mcp.json, select Refresh in Settings → MCP, or ask the agent to refresh MCP tools. Newly connected capabilities then appear in the current conversation, typically with names beginning with mcp_*.

Connected servers make their tools available to the agent. Turning on Shared with everyone makes a server available beyond your private instance, using a sharing model similar to skills and app integrations.

Inside the desktop container, you can also inspect MCP through the endos-mcp command-line interface described in Native APIs and CLIs.