Loading ToolBench
Unified macOS automation MCP server with 17 atomic tools for UI interaction, app control, and system automation via accessibility APIs, Chrome DevTools Protocol, and AppleScript
Argus provides 16 well-named tools with complete input schemas and descriptions. Naming follows verb_noun convention (argus_click, argus_type, argus_find). Descriptions are concise (50 - 150 chars) and action-oriented. However, parameter descriptions are sparse or missing entirely, most params lack explanation of constraints, ranges, or expected formats. Output schemas are not documented; LLMs cannot predict response structure. Error handling is implicit (no recovery guidance in descriptions). Security-sensitive tools (argus_exec_apple_script, argus_type on secure fields) have basic safeguards but lack audit trail documentation. Composition is strong: tools chain well (find → click → verify), and responses include coordinates for downstream calls.
Cascade resolve + click + verify. Returns {clicked, x, y, source, confidence, verified:{ok,reason}, attempts, ms}. Updates the learned-selector cache on success.
Diagnostics. Returns {status: READY|DEGRADED|BROKEN, summary, deps, surface, dotenv}. Pass full=true for the verbose blob (ax/ocr/cdp/vision/cgevent details).
Run arbitrary AppleScript. Escape hatch — sandboxed via safety.is_dangerous(). Refuses on rm -rf / fork bombs / disk-erase / admin escalation. Pass allow_dangerous=true (audit-logged) to override.
Resolve target to coords WITHOUT clicking. Returns {x,y,bbox?,source,confidence,scope,attempts,ms}. Cascade: cache → AX → CDP-DOM → OCR → vision.
Return ALL matches for target (not just the best). Useful for duplicates, e.g. 'pick the second Sign In button'. Each hit has {x, y, bbox, source, confidence, text?}.
Special key (Return, Tab, Escape, ArrowUp/Down/Left/Right, Backspace) with optional modifiers ('cmd'|'shift+cmd'|'ctrl'|'alt').
Parameter descriptions missing or minimal. Most params (e.g., 'combo' in argus_send_keys, 'modifiers' in argus_key, 'action' in argus_session) lack explanation of valid values, constraints, or expected format. LLMs cannot infer whether 'combo' accepts 'cmd+shift+t' or 'command+shift+t'.
Output schemas not documented. Tool descriptions state what is returned (e.g., '{x,y,bbox?,source,confidence}' for argus_find) but not as formal JSON Schema. LLMs cannot reliably extract or chain results without explicit response structure.
Error handling lacks recovery guidance. No descriptions explain what to do if a tool fails (e.g., 'If argus_click fails with confidence < min_confidence, try argus_find_all to see all matches'). Agents cannot self-correct.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
Launch / focus a macOS app by name.
Paste text into focused field via clipboard (much faster than typing for long text; handles Unicode/IME natively). Refuses on AXSecureTextField unless force_secure=true. Restores prior clipboard contents.
Quit a macOS app gracefully.
Scroll up/down/left/right at cursor (or at coords if provided).
Screenshot of the current screen → base64 PNG.
Composite keyboard shortcut as single string, e.g. 'cmd+shift+t' or 'ctrl+alt+delete'. Maps the last token to the key, the rest to modifiers.
Browser session jar (cookies + storage). action ∈ {save, load, clear, list, detect_login}.
Where am I? Returns {surface: 'browser'|'webview'|'native', app: {bundle_id,name,pid}, host?}
Type into focused field. Refuses on AXSecureTextField unless force_secure=true.
Poll until target appears (or condition met) within timeout. Returns the resolved hit, or {timed_out: true}. Use to eliminate sleep() guesswork.
Numeric parameter constraints not specified. argus_see (max_dim), argus_find_all (limit, min_confidence), argus_wait_for (timeout_s, poll_s) lack min/max bounds. LLMs may pass absurd values (max_dim=999999, timeout_s=0.001).
argus_session action parameter is a free-form string with no enum constraint. Description lists valid actions (save, load, clear, list, detect_login) but does not enforce them. LLMs may pass invalid actions like 'delete' or 'export'.