Connecting MCP servers
The host can connect MCP servers and the agent gets their tools, whatever model it runs on.
The host can connect MCP servers and the agent gets their tools, whatever model it runs on.
Servers live in .mcp.json in the project or ~/.plural/mcp.json for every session, in the
same mcpServers format Claude Code and Cursor use, so a config can be copied over as is.
Commands
| Command | What it does |
|---|---|
/mcp | The servers, their status and tool count; ↑↓ chooses, enter turns one on or off, ctrl+r restarts it. + add a server… opens a wizard with popular servers (Context7, Playwright, GitHub, Notion, Filesystem, Fetch, Git, Memory, Time) or your own command or URL, and asks only for what that server needs |
/mcp import | Takes the servers already set up for Claude Code (~/.claude.json) and Cursor (~/.cursor/mcp.json) |
/mcp add <name> <command…> | Launches a local server, for example /mcp add plane npx -y @makeplane/plane-mcp-server |
/mcp add <name> <url> | Connects a remote server over HTTP |
/mcp on|off|restart|rm <name> | Turns a server on or off, restarts or removes it |
--global on add saves the server for every project. Environment variables and HTTP headers
such as tokens go in the JSON file; they never reach the room and are scrubbed from tool output.
Remote servers that sign in with OAuth are not supported yet.
{
"mcpServers": {
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "…" }
}
}
}Every call still waits for the host
Every MCP tool call waits for the host's approval like a command, and always allows that
tool for the rest of the session. A run with a participant's prompt still asks in auto mode,
and plan mode refuses MCP calls. A server writes its stderr to
~/.plural/state/logs/mcp-<name>.log.