llx as an MCP server

llx mcp serves llx's window to another agent (any client that speaks MCP over stdio): native browser pages and terminal sessions as tabs of one window you watch, grouped by project, with the same actions, kinds and snapshots the llx agent has.

The other agent's model does the thinking, so no chat model is used. A goal on a tab is carried out by llx's decision model, which needs an OpenRouter key (OPENROUTER_API_KEY, or one saved with /login). Without one, everything but goal works.

Connect#

Any MCP client takes a server whose command is llx with the argument mcp (install it with npm install -g @localixai/llx):

{
  "mcpServers": {
    "llx": { "command": "llx", "args": ["mcp"] }
  }
}

With nothing installed, the command is npx with the arguments -y @localixai/llx mcp: npx fetches the llx binary for your platform.

The goal of a tab needs an OpenRouter key: save one with /login in llx, or set OPENROUTER_API_KEY in the server's environment. Without one, everything but goal works.

Wrappers for particular agents live outside this repository; llx mcp is the interface they use.

How it runs#

• It runs in the daemon (started if none runs), so its tabs share the one window with your llx sessions, under the project the agent started it in; its terminals run there, with the agent's environment.
• The other agent approves each call (its own permission prompt): llx asks nothing more for them.
• A cancelled call is stopped. When the agent closes the server, its terminals end and the browser tabs it left open close.
• Comments the person sends from its tabs come back with the agent's next window call, after that call's report, under ### What the person sent from the window. A wrapper can ask for them with a quiet list to have them arrive without the agent calling.
• Unix only. The headless Linux build has no browser, so its window has terminal tabs only.

The rules behind it are W9 and K7 in ARCHITECTURE.md.