List all surfaces (terminal/browser panes) across workspaces

Create a new split pane (terminal or browser)

Send text input to a terminal surface. When sending commands to another Claude session, press_enter can be unreliable — for critical inputs, use send_input without press_enter, then call send_key with key 'return' separately.

Send a key press to a terminal surface. Use this after send_input to reliably submit commands — especially when targeting interactive programs like Claude sessions.

Read the current screen content of a terminal surface

Rename a surface tab

Set a sidebar status key-value pair

Set sidebar progress indicator (0.0 to 1.0)

Close a surface (terminal or browser pane)

Interact with a browser surface (open, navigate, snapshot, click, type, eval, wait)

Spawn an AI agent in a new terminal surface. Returns immediately — use wait_for to block until ready.

Block until an agent reaches a target state (ready, done, error). Checks retroactively first.

Block until ALL agents reach target state OR any agent errors (fail-fast with partial results).

Get the full state of an agent including cli_session_id for resume.

List all agents with optional filters by state, repo, or model.

Stop an agent gracefully (Ctrl+C) or forcefully (kill process).

Send text input to an agent. Agent must be in ready or idle state.

Extract structured output from an agent's terminal between delimiter markers (e.g., REVIEW_OUTPUT_START / REVIEW_OUTPUT_END). Returns the content between the markers, or null if not found.

Send a message to an agent, or perform an agent action (interrupt, model switch, resume, skill, usage). If the agent is alive, sends directly. If not found, returns an error — use spawn_agent first.

Stop one or more agents. Target can be a single agent ID, an array of IDs, or 'all'.