Loading ToolBench
Mac Local Vision (Zero-Token OCR / E2E targeting / on-device vision) — MCP server providing local vision capabilities over stdio and HTTP transports
Strong schema definitions and descriptions across 11 tools. All tools have explicit input schemas with typed parameters and clear descriptions (avg 150+ chars). Tool names are action-oriented (ocr, find, classify, ask). However, output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal; no recovery guidance or actionable error messages. Missing tool annotations (readOnlyHint/destructiveHint). Some parameter descriptions could be more prescriptive about constraints (e.g., minConfidence range 0..1 is stated but not enforced in schema). The 'ask' tool requires macOS 27 Beta, this critical dependency is buried in description, not surfaced as a prerequisite check.
On-device multimodal reasoning (Beta — needs macOS 27 (Beta) + an Apple-Intelligence-eligible Mac). Accepts an image and a prompt, optionally with a JSON Schema for Guided Generation.
Scan QR/1D/2D barcodes from an image or PDF.
Tag an image with a 1,303-label taxonomy classification.
Report which modes are available on this machine.
Find a document's 4 corners in an image or PDF.
Structured OCR extracting title/paragraphs/tables/lists from an image or PDF.
Locate a specific word/phrase on a screenshot and return the exact pixel center to click (x,y) plus its bounding box — for fast E2E/UI targeting. ALWAYS check the `found` field: it is false when the target is absent. When `approximate: true` is present, the box is the whole text line (click point is line-center, not word-tight). Coordinate convention: top-left origin
Output schemas not documented. LLMs cannot predict response structure (fields, types, nesting). Forces agents to guess at downstream field names and types.
No error handling guidance. Errors lack recovery hints (e.g., 'File not found. Check path or use base64 data parameter'). Agents cannot self-correct.
Tool annotations missing. No readOnlyHint/destructiveHint/idempotentHint. Agents cannot determine which tools are safe to retry or have side effects.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 65 | 2026-07-28+ | v2 |
Generate a scannable QR code PNG from text (CoreImage only).
Zero-token OCR: read all text from an image or PDF locally on the Mac NPU (no cloud, no vision tokens spent). Returns the full text plus per-line entries. Coordinates are opt-in to stay token-lean: set boxes=true for per-line pixel boxes, words=true for per-word boxes. Coordinate convention: x,y = click center; left/top/width/height = bounding box; top-left origin; physical pixels. Recognition languages auto-detect from the system locale (override via languages). Use this to read a whole screen; use `find` to target one specific word. Remote callers (non-Mac nodes): send the image as base64 in the `data` field instead of `path`, or use HTTP method QUERY (RFC 10008) — e.g. curl -X QUERY "http://host:9090/ocr?languages=ko-KR&format=json" --data-binary @file.jpg where the body is raw file bytes and options go in the query string (make-qr takes UTF-8 text as its body).
Scan for QR codes only from an image or PDF.
Flatten a photographed document (Vision detect + CoreImage).
'ask' tool has critical platform dependency (macOS 27 Beta + Apple Intelligence). Not surfaced as a prerequisite or availability check. Agents may invoke on incompatible systems.
Parameter constraints (minConfidence 0..1, page ≥1, scale ≥1) stated in descriptions but not enforced in JSON Schema (no minimum/maximum fields). Agents may pass invalid values.