| name | computer-use |
| description | Use when a user asks to inspect, navigate, or operate an explicitly allowed macOS application through the blade-computer-use MCP tools. |
Computer Use
Operate one allowed macOS app at a time. Treat every observation revision as a single-use capability for exactly one action.
Workflow
- Call
list_apps and identify the target bundle identifier.
- Call
observe for that app. Use element_index when a suitable AX node exists; use window-relative coordinates only as a fallback.
- Confirm immediately before impactful actions such as sending, submitting, purchasing, deleting, publishing, or changing permissions.
- Perform one
click, type_text, press_key, or scroll with the revision returned by that observation.
- Observe after every action. Never reuse a revision or assume the interface remained unchanged.
If an action reports a stale or consumed revision, observe again and reassess the current interface. Do not replay the old action blindly.
Tools
| Tool | Purpose |
|---|
list_apps | List running GUI apps and bundle identifiers. |
observe | Return the current AX tree, revision, and optional screenshot. |
click | Click an AX element or window-relative coordinate. |
type_text | Type Unicode text into the focused non-secure field. |
press_key | Press a supported key with optional modifiers. |
scroll | Scroll at an element, coordinate, or window center. |
Safety
- Stay within the app the user named and the configured
CUA_ALLOWED_APPS policy.
- Explain when Accessibility or Screen Recording permission is missing; do not weaken the policy to bypass it.
- Never target
loginwindow, SecurityAgent, password fields, or other secure input surfaces.
- Pause when the visible state differs from the user's request or when physical user input changes the target.
- Prefer a purpose-built API or CLI when it can complete the task with less UI access.
Common Mistakes
- Reusing a revision: observe again before retrying.
- Guessing screen coordinates from an old screenshot: use the current AX element index first.
- Chaining several actions from one observation: one observation permits one action only.
- Treating a successful input event as task completion: verify the resulting state with
observe.
