Exposes the Hailo-8L accelerator as 14 MCP tools for computer vision tasks including face detection, object detection, OCR, pose estimation, segmentation, depth estimation, and text/image embeddings.
14 tools with complete input schemas and descriptions, but quality is uneven. Naming is strong (verb_noun pattern, clear intent). Descriptions are detailed but verbose (200-400 chars, above the 194-char baseline), mixing implementation details with user-facing guidance. Parameters are well-typed with defaults, but lack constraint documentation (enums, ranges, patterns). Output schemas are documented in descriptions but not formally specified. Error handling is mentioned (structured error dicts) but recovery guidance is minimal. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Security: file paths are user-provided (path traversal risk), no input validation visible, no audit logging.
Preview-grade relative depth map via Depth-Anything-V2 (ViT-S) on Hailo-8L. Writes a 224×224 8-bit PNG (bright = near) to out_path, or next to the input as <name>.depth.png. Returns {"depth_path", "min", "max", "mean"}. Good for shot analysis and parallax previews; not a production depth pass.
Compute a 512-d TinyCLIP image embedding on Hailo-8L (for similarity / clustering).
Brighten an under-exposed frame via Zero-DCE on Hailo-8L (preview-grade). Writes the enhanced image at the original resolution to out_path, or next to the input as <name>.enhanced.png. Returns {"enhanced_path"}.
Detect faces in a JPEG/PNG via SCrFD-2.5G on Hailo-8L. Returns {"faces": [{x,y,w,h,score,kps}, ...], "count": N} on success, or a structured {"error": True, ...} dict while disabled or on device error. kps = the 5 landmarks (left eye, right eye, nose, mouth-L, mouth-R) in image pixels — pass a face straight into hailo_face_embed for identity.
Identity vectors for every face in an image (SCrFD detect → ArcFace). For each face returns the box plus a 512-d L2-normalised ArcFace embedding. Cosine similarity between two embeddings is the identity score: the same person scores ≈0.5+ across shots, different people ≈0.3-. Use it to cluster who appears where across a whole project without a single cloud call. Faces are processed strongest-score first, up to max_faces.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools like hailo_depth, hailo_enhance_low_light, hailo_segment perform writes but lack destructiveHint. Agents cannot distinguish safe reads from state-modifying operations.
Descriptions for hailo_embed and hailo_text_embed are under 50 chars ('Compute a 512-d TinyCLIP image embedding...').
No input validation or constraint documentation. score_threshold parameters lack min/max bounds (e.g., 0.0 - 1.0). max_faces lacks upper bound. LLMs can pass invalid values (negative scores, 1000 faces) without guidance. No enums for categorical choices.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Detect the 80 COCO object classes (person, car, dog, laptop, ...) via YOLOv8s with on-chip NMS on Hailo-8L. Returns {"objects": [{label, class_id, x, y, w, h, score}, ...], "count": N}, sorted by score. Boxes are in the image's own pixel space.
Extract text from a thumbnail/frame via PaddleOCR v5 (det + rec) on Hailo-8L. Returns {"text": str, "char_count": int, "boxes": [[x,y,w,h], ...]}.
Re-identification vectors for every PERSON in an image (YOLOv8s → OSNet). Unlike face identity this works with no visible face — it keys on clothing and body shape — so it tracks the same person across shots and angles. Returns per person: the box plus a 512-d L2-normalised embedding; cosine similarity between two embeddings is the match score.
Human pose estimation via YOLOv8s-pose on Hailo-8L (host-decoded raw head). Returns {"people": [{x,y,w,h,score,keypoints:{nose:{x,y,score},...}}], "count": N} — 17 COCO keypoints per person, in the image's own pixel space.
Instance segmentation via YOLOv8s-seg or FastSAM-s on Hailo-8L. Returns instance masks as an id-map PNG (one color per instance) plus boxes and class labels.
Device + runtime state, safe with driver broken.
Compute a 512-d TinyCLIP/SigLIP2 text embedding on Hailo-8L (text tower on-NPU).
Transcribe audio via Whisper-base encoder+decoder HEFs on Hailo-8L (5 s chunks).
Free-text zero-shot classification: rank image against text labels via TinyCLIP/SigLIP2 (both towers on-NPU). Returns ranked similarities between the image and each label.
File path parameters (image_path, audio_path, out_path) are user-provided strings with no validation. Path traversal attacks possible (e.g., '../../../etc/passwd'). No sanitization visible in source. Descriptions do not warn about path constraints.
Error handling returns structured dicts but lacks recovery guidance. Descriptions mention 'structured error dicts' but do not specify error categories (retryable, user-fixable, fatal) or actionable next steps. LLMs cannot distinguish 'retry' from 'ask user' from 'give up'.