Skip to content

Latest commit

 

History

History
84 lines (64 loc) · 7.14 KB

File metadata and controls

84 lines (64 loc) · 7.14 KB
summary Cheat sheet for every Peekaboo CLI command grouped by category.
read_when
learning what each CLI subcommand does
mapping agent tools to direct CLI usage

CLI Command Reference

This source-tree reference covers all 34 root commands in the upcoming v4 registry and is checked against the built binary's --help output. Use peekaboo <command> --help for every option and peekaboo tools for the separate MCP/agent tool catalog.

Core commands

Command Purpose / subcommands
bridge Inspect Bridge connectivity; status is the default subcommand.
capture action, live, and video capture workflows.
clean Remove snapshot cache data, with --dry-run support.
completions Generate zsh, bash, or fish completion scripts.
config Configuration plus credential and provider subcommand trees.
daemon run, start, status, and stop the headless daemon.
learn Print the agent guide, tool catalog, and live command signatures.
permissions status, grant, or request <kind>.
screen list connected displays.
tools list MCP tools or describe <name> for one schema.

Interaction commands

Command Purpose
action Invoke a named accessibility action.
click Click an element/query or --at x,y.
drag Drag between element IDs or coordinates using --from and --to.
move Move the physical pointer to --on or --at.
paste Paste current clipboard content or atomically set, paste, and restore.
press Press xdotool-style chords or chord sequences.
scroll Scroll by direction on an element or an exact-window --at point.
select-text Select literal text or place a caret before/after it without focusing or typing.
set-value Set an accessibility element value directly.
type Type text; standalone keys and chords belong to press.

System commands

Command Purpose / subcommands
app focus, hide, launch, list, quit, relaunch, switch, and unhide.
clipboard get, set, clear, save, and restore.
dialog click, dismiss, file, input, and list.
dock hide, launch, list, right-click, and show.
menu click or list application menus.
menubar click or list menu-bar status items.
space list, move-window, and switch Spaces.
visualizer Exercise the agent cursor, input HUD, and capture indicators.
window close, focus, list, maximize, minimize, move, resize, restore, and set-bounds.

Vision, AI, and MCP commands

Command Purpose / subcommands
see Capture pixels and element maps; use --tree, --no-screenshot, or --no-elements to select the observation shape.
verify Poll stable window/element predicates; results are satisfied, unsatisfied, or unknown.
agent run, resume, sessions, and chat; run is the default.
browser Control Chrome page content through the browser MCP tool.
mcp Start the MCP server; serve is the default subcommand.

Shared grammar

Durations accept bare milliseconds, ms, or s: 500, 500ms, 2s, and 1.5s are equivalent forms. Coordinate input is --at x,y; with an app/window target it is target-relative unless --global is present. Modifier lists use comma-separated values such as cmd,shift.

Command-deadline errors retain fractional durations: 200ms is reported as 200 milliseconds, and 1.5s as 1.5 seconds. Positive subsecond durations are rounded to the nearest millisecond, with a minimum displayed value of one millisecond. This formatting does not change the actual deadline or TIMEOUT error code.

Interaction commands share foreground/focus controls where relevant. Background delivery is the default when Peekaboo can resolve an exact process target; physical pointer gestures and intentional global input require --foreground.

JSON result envelope

Pass --json (or the Commander-provided --json-output alias) for one stable result shape. Every response has success, data (null when unavailable), optional error, and debug_logs. Failed responses exit nonzero and include error.code, error.message, and an actionable error.hint when the command already knows the next step.

success: true and exit 0 indicate command-level success, not confirmed receiver effect. Human CLI output and MCP action summaries distinguish unverified dispatch from confirmed application change; inspect the recorded safety fields and observe the target before retrying instead of automatically replaying an unverified action.

Once the command path identifies an action request, its result also includes a top-level effect: confirmed when existing AX/readback verification proves the result, partial for a partly completed multi-step action, unverifiable when input was dispatched without an application-level signal, suspected_noop when a post-check found no change, or refused when a safety gate prevented dispatch. Commander parse/bind failures for recognized action commands report effect: refused; read-only commands omit effect. MCP action tools expose the same canonical fields in result metadata.

When an interaction, window mutation, or application lifecycle command returns a native action receipt, JSON also includes a top-level outcome object. This includes click, type, scroll, press, action, set-value, select-text, background window geometry/lifecycle operations, and application launch, relaunch, quit, hide, unhide, focus, and switch. The object is the validated canonical projection: state, route, delivery, evidence, dispatch state/count, retry safety, escalation, refusal reason, and the derived mutation_dispatched, retry_safe, and requires_fresh_observation compatibility fields. The legacy top-level effect and failure safety fields derive from that same object. Read-only commands, older hosts, and actions without native receipts omit outcome rather than fabricating a receipt.

For a completed action from a legacy provider without a canonical outcome, human output explicitly says “request completed; receiver effect was not reported.” This wording does not fabricate an outcome or receipt, infer confirmation, or change the existing JSON, exit status, or retry rules.