llui-agent

MCP bridge for the LLui Agent Protocol. Install once into your LLM client; paste the connect snippet from any LLui app to bind the conversation to that app.

Install (Claude Desktop)

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or equivalent on your OS:

{
  "mcpServers": {
    "llui": {
      "command": "npx",
      "args": ["-y", "llui-agent"]
    }
  }
}

Restart Claude Desktop. The 11 LLui tools (connect_session, disconnect_session, describe_app, get_state, list_actions, send_message, get_confirm_result, wait_for_change, query_dom, describe_visible_content, describe_context) now appear. Desktop exposes the bundled llui-connect MCP prompt as a slash command — see "Slash shortcuts" below.

Install (Claude Code CLI)

claude mcp add --transport stdio llui -- npx -y llui-agent

For local development against unpublished bridge code, point at the built CLI instead:

claude mcp add --transport stdio llui -- node /absolute/path/to/llui/packages/agent-bridge/dist/cli.js

Run /mcp inside CC to confirm the server connected (or start a new session). The same 11 tools become available.

If you run CC in auto mode (permissions.defaultMode: "auto" in ~/.claude/settings.json), the auto-classifier silently rejects unrecognized MCP tools the first time they're called — Claude reports "tool was rejected" but no UI prompt is shown. Add the bridge's tools to your allowlist once so subsequent calls go through:

// ~/.claude/settings.json
{
  "permissions": {
    "allow": [
      "mcp__llui__*", // replace `llui` with the name you used in `claude mcp add`
    ],
  },
}

Users on defaultMode: "default" or "ask" instead get a permission prompt on the first call and don't need this allowlist entry.

Use

Open any LLui app built with @llui/agent/client. Click "Connect with Claude" in the app and copy the generated snippet. Paste it into Claude — the snippet is a natural-language instruction containing the URL and token. Claude reads it and calls connect_session to bind. The same snippet works in Desktop and CC.

Each Claude chat is bound to ONE LLui app at a time. To switch, ask Claude to call disconnect_session and paste a new snippet.

Slash shortcuts (optional)

The bridge registers an MCP prompt named llui-connect. Both clients surface it as a slash command, but the namespacing differs:

| Client | Shortcut | | --------------- | ------------------------------------------------- | | Claude Desktop | /llui-connect <url> <token> | | Claude Code CLI | /mcp__<server-name>__llui-connect <url> <token> |

The <server-name> in CC is whatever you passed to claude mcp add — llui if you used the command above. Power-user shortcut only; the natural-language snippet from the app works the same in either client and doesn't depend on the server-name choice.

How it works

1. Your LLui app mints a per-browser-session token and renders a connect snippet — a one-line instruction containing the LAP URL and the bearer token.
2. You paste into Claude — Claude reads the snippet, calls connect_session, and the bridge records {url, token} for this chat.
3. The bridge pings POST {url}/describe to validate and cache the app's schema.
4. Subsequent Claude tool calls (get_state, send_message, etc.) forward to {url}/<path> with your token as a Bearer.
5. Sensitive actions (@requiresConfirm in the app's code) route through a confirmation prompt that only the user can approve.