The MCP server

Most terminals are something agents run inside. Novalis is also something agents can fly: the running window exposes a Model Context Protocol server with 38 tools.

Start it

# Standalone / headless
novalis --mcp-server

# Or against the live window
NOVALIS_LIVE_MCP=1 open -a Novalis

Or flip it in Settings → MCP: the toggle hot-applies to the running window.

Register it with your agent of choice (Claude Code shown):

claude mcp add novalis -- <path-to-bridge>

From then on, an agent session holds your terminal as native tools.

What agents can do

Category	Tools (selection)
Workspaces & tabs	create_workspace, open_tab (workspace-targeted), close_tab, select_tab, list_tabs
Panes	list_panes (split tree, per-pane cwd, alt-screen state), split_pane, focus_pane, close_pane
Reading	read_pane (scrollback), list_blocks, list_notifications
Synchronization	wait_for_output (regex over live output), wait_for_idle (grid quiet for N ms)
Agents	spawn_agent_pane; a pane anywhere running an agent CLI with a prompt; agents spawning agents in one call
Signals	notify; post into the attributed notification rail

The philosophy

The MCP surface rides the same intent pipeline as the UI; an agent splitting a pane goes through exactly the code path your ⌘D does. Nothing is a simulation layer, and everything an agent does is visible in the window, in real panes, in real blocks.

Write-capable verbs respect the same approval architecture as agent mode; the human owns the gate.