Install any skill in seconds. Free to start, no credit card required.
Get Started Free →GUI control for macOS apps via mediar-ai's mcp-server-macos-use. Click, type, scroll, key-press, open apps — driven by accessibility tree, works in non-interactive Claude Code mode. Use this for any Sutando task that needs to drive another macOS application (Safari, Zoom, Mail, Finder, etc.).
.claude/skills/sonichi-macos-use/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-06 | ✗→✓ | ▲ Improved | -49% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -2% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -15% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -14% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -20% | 0% |
Drive macOS applications from Claude Code via mediar-ai's mcp-server-macos-use. A Swift MCP server that wraps the macOS Accessibility API. Unlike Claude's built-in computer-use, this works in non-interactive mode (which is how Sutando's proactive loop and task bridge run), does not hold a machine-wide lock, and does not require a Pro/Max subscription.
src/inline-tools.tsPrefer this skill over:
bash src/screen-capture.sh — that captures screenshots; macos-use actually interactstell application blocks — more reliable, better error handlingcliclick — lower-level, no accessibility contextcomputer-use — that mode requires interactive sessions and holds a lock that contends with Sutando's own loopAfter install, these appear as mcp__macos-use__* in Claude Code:
| Tool | Parameters | Purpose | |------|------------|---------| | open_application_and_traverse | identifier (name/bundle ID/path) | Launch or activate an app, return its a11y tree | | click_and_traverse | pid, x, y | Click at coordinates in a target app, return updated tree | | type_and_traverse | pid, text | Type into the frontmost element | | press_key_and_traverse | pid, key | Press a named key (Return, Tab, Escape, arrows, ...) | | scroll_and_traverse | pid, direction, amount | Scroll in a direction | | refresh_traversal | pid | Re-read the a11y tree without acting |
Every tool returns an accessibility-tree snapshot of the target app — structured UI elements with roles, titles, positions, and identifiers. No pixels. Model reasons over the tree, not over screenshots.
Two steps, one-time:
bash# 1. Build the Swift binary (~35s) bash skills/macos-use/scripts/build.sh # 2. Register with Claude Code's MCP config (writes ~/.claude.json) bash skills/macos-use/scripts/install-mcp.sh # 3. Grant Accessibility permission # System Settings → Privacy & Security → Accessibility # Click +, navigate to ~/.macos-use-mcp/.build/release/mcp-server-macos-use, enable.
Restart Claude Code after install for the MCP tools to appear.
swift-sdk transitive dep has data-race errors that Swift 6.3+ strict-concurrency trips on. build.sh uses -Xswiftc -swift-version -Xswiftc 5 as a workaround. When upstream fixes this, remove the flag.screen-capture.sh + Claude vision.sutando-memory.git (binaries are machine-specific). Run build.sh + install-mcp.sh on Mac Mini and MacBook separately.After install + restart:
Sutando, open Safari and navigate to https://github.com/sonichi/sutandoYou should see Claude invoke mcp__macos-use__open_application_and_traverse with identifier: "Safari", then type_and_traverse into the URL bar, then press_key_and_traverse with Return.
notes/issue-65-computer-use-research.md| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-22 | pass→pass | 16,300 | 5,661 | -65% | 1 | 1 | 0% | 2,444 | 1,928 | -21% | 0 | 0 | — |
case-06 | fail→pass | 16,010 | 1,575 | -90% | 1 | 1 | 0% | 2,619 | 1,340 | -49% | 0 | 0 | — |
case-01 | fail→fail | 13,456 | 15,067 | +12% | 1 | 1 | 0% | 2,399 | 1,811 | -25% | 0 | 0 | — |
case-02 | fail→fail | 11,540 | 8,023 | -30% | 1 | 1 | 0% | 2,085 | 1,459 | -30% | 0 | 0 | — |
case-03 | fail→fail | 13,740 | 8,234 | -40% | 1 | 1 | 0% | 2,298 | 2,509 | +9% | 0 | 0 | — |
case-04 | pass→pass | 11,669 | 2,961 | -75% | 1 | 1 | 0% | 1,935 | 1,581 | -18% | 0 | 0 | — |
case-05 | fail→pass | 10,324 | 2,675 | -74% | 1 | 1 | 0% | 1,605 | 1,568 | -2% | 0 | 0 | — |
case-07 | fail→fail | 7,442 | 3,958 | -47% | 1 | 1 | 0% | 1,240 | 1,697 | +37% | 0 | 0 | — |
case-08 | fail→fail | 7,512 | 1,977 | -74% | 1 | 1 | 0% | 1,228 | 1,420 | +16% | 0 | 0 | — |
case-09 | fail→pass | 19,215 | 8,468 | -56% | 1 | 1 | 0% | 2,751 | 2,347 | -15% | 0 | 0 | — |
case-10 | fail→pass | 11,417 | 2,553 | -78% | 1 | 1 | 0% | 1,696 | 1,458 | -14% | 0 | 0 | — |
case-11 | pass→pass | 6,998 | 1,318 | -81% | 1 | 1 | 0% | 1,089 | 1,256 | +15% | 0 | 0 | — |
case-12 | fail→pass | 15,118 | 3,853 | -75% | 1 | 1 | 0% | 2,052 | 1,651 | -20% | 0 | 0 | — |
case-13 | pass→pass | 10,355 | 2,610 | -75% | 1 | 1 | 0% | 1,512 | 1,447 | -4% | 0 | 0 | — |
case-14 | pass→pass | 11,520 | 2,844 | -75% | 1 | 1 | 0% | 1,458 | 1,513 | +4% | 0 | 0 | — |
case-15 | pass→pass | 7,221 | 2,301 | -68% | 1 | 1 | 0% | 1,090 | 1,397 | +28% | 0 | 0 | — |
case-21 | fail→pass | 13,666 | 6,310 | -54% | 1 | 1 | 0% | 2,124 | 2,152 | +1% | 0 | 0 | — |
case-16 | fail→pass | 13,623 | 3,108 | -77% | 1 | 1 | 0% | 1,890 | 1,375 | -27% | 0 | 0 | — |
case-17 | fail→pass | 14,665 | 1,458 | -90% | 1 | 1 | 0% | 2,057 | 1,291 | -37% | 0 | 0 | — |
case-18 | fail→pass | 17,530 | 1,942 | -89% | 1 | 1 | 0% | 2,558 | 1,339 | -48% | 0 | 0 | — |
case-19 | fail→pass | 5,661 | 1,706 | -70% | 1 | 1 | 0% | 896 | 1,346 | +50% | 0 | 0 | — |
case-20 | pass→pass | 14,749 | 7,144 | -52% | 1 | 1 | 0% | 2,268 | 2,242 | -1% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 20 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +45 percentage points is the difference between those two pass rates over the 20 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.