Has Anonymizer

---
name: has-anonymizer
description: "HaS (Hide and Seek) on-device text and image anonymization. Text: 8 languages (zh/en/fr/de/es/pt/ja/ko), open-set entity types. Image: 21 privacy categories (face, fingerprint, ID card, passport, license plate, etc.). Use when: (1) anonymizing text before sending to cloud LLMs then restoring the response, (2) anonymizing documents, code, emails, or messages before sharing, (3) scanning text or images for sensitive content, (4) anonymizing logs before handing to ops/support, (5) masking faces/IDs/plates in photos before publishing or sharing."
metadata:
  {
    "openclaw":
      {
        "emoji": "🔒",
        "requires": {
          "bins": ["llama-server", "uv"],
          "env": {
            "HAS_TEXT_MODEL_PATH": {
              "description": "Absolute path to the HaS text GGUF model (default: ~/.openclaw/tools/has-anonymizer/models/has_text_model.gguf)",
              "required": false
            },
            "HAS_IMAGE_MODEL": {
              "description": "Absolute path to the HaS image YOLO model (default: ~/.openclaw/tools/has-anonymizer/models/sensitive_seg_best.pt)",
              "required": false
            },
            "HAS_TEXT_MAX_PARALLEL_REQUESTS": {
              "description": "Maximum parallel inference requests for scan/hide/seek batch operations (default: 4, max: 4)",
              "required": false
            }
          }
        },
        "install":
          [
            {
              "id": "brew-uv",
              "kind": "brew",
              "os": ["darwin"],
              "formula": "uv",
              "bins": ["uv"],
              "label": "Install uv (brew)",
            },
            {
              "id": "brew-llama",
              "kind": "brew",
              "os": ["darwin"],
              "formula": "llama.cpp",
              "bins": ["llama-server"],
              "label": "Install llama.cpp (brew)",
            },
            {
              "id": "download-model",
              "kind": "download",
              "url": "https://huggingface.co/xuanwulab/HaS_Text_0209_0.6B_Q8/resolve/main/has_text_model.gguf",
              "targetDir": "models",
              "label": "Download HaS text model Q8 (639 MB)",
            },
            {
              "id": "download-image-model",
              "kind": "download",
              "url": "https://huggingface.co/xuanwulab/HaS_Image_0209_FP32/resolve/main/sensitive_seg_best.pt",
              "targetDir": "models",
              "label": "Download HaS image model FP32 (119 MB)",
            },
          ],
      },
  }
---

# HaS Privacy

HaS (Hide and Seek) is an on-device privacy protection tool. It provides **text** and **image** anonymization capabilities, both running entirely on-device.

- **Text anonymization** (has-text): Powered by a 0.6B privacy model, supports 8 languages with open-set entity types for anonymization and restoration
- **Image anonymization** (has-image): Powered by a YOLO11 segmentation model, supports pixel-level detection and masking of 21 privacy categories

## Agent Decision Guidelines

- **First introduction**: When users encounter HaS for the first time, demonstrate value through real-world scenarios rather than listing commands. Examples: anonymize contracts/resumes before sharing safely, anonymize before sending to cloud LLMs then restore the response, auto-mask faces/IDs/license plates and 20 other privacy categories in photos before publishing, scan workspace for privacy leak risks, anonymize logs before handing to ops/support
- **Scanning workspace/directory**: Use has-text scan for text files and has-image scan for image files simultaneously, then provide a consolidated report
- **Non-plaintext formats**: has-text only processes plaintext. For PDFs, Word documents, scanned images, etc., first convert to text using other available tools before processing
- **Text in images**: has-image covers most text-in-image scenarios by masking all 21 visual carriers (screens, paper, sticky notes, shipping labels, etc.) as a whole. For further recognition of text content in images, use OCR to extract text first, then run has-text scan for additional detection
- **Never delete original files**: Anonymization operations should output to new files, **never overwrite or delete the original files**. Image anonymization is irreversible; text anonymization can be restored but the original file should still be preserved as backup
- **Proactively inform about configurable options**: At appropriate moments, inform users about the following options and help them configure interactively:
  - **Text**: `--types` can specify any entity type (names, addresses, phone numbers, etc.), not limited to predefined types
  - **Image**: `--types` can specify which categories to mask (e.g., only faces, or only license plates), defaults to all 21 categories
  - **Masking method**: `--method` supports mosaic (default), blur, and solid color fill
  - **Masking strength**: `--strength` adjusts mosaic block size or blur intensity (default 15)
- **Post-scan report**: After scanning a workspace/directory, generate a consolidated privacy check report including:
  - Total files scanned (text and image counts separately)
  - Number and location of each type of sensitive content found
  - Risk level assessment (flag high-sensitivity items such as ID numbers, faces, etc.)
  - Recommended next steps (e.g., "Would you like to anonymize the above files?")
- **Report elapsed time after completion**: After task completion, report processing time to the user so they can perceive on-device inference performance. For single tasks, report individual time (e.g., "On-device inference complete, took 0.3s"); for batch tasks, report a summary (e.g., "Processed 12 texts + 8 images, total time 2.4s"). Do not display technical metrics like tok/s

---

# Part 1: Text Anonymization (has-text)

## Core Concepts

### Three-Level Semantic Tags

Tag format after anonymization: `<EntityType[ID].Category.Attribute>`

- **EntityType**: e.g., person name, address, organization
- **[ID]**: Sequential number for entities of the same type. After coreference resolution, the same entity shares the same number — "CloudGenius Inc.", "CloudGenius", and its Chinese equivalent all map to `<Organization[1].Company.CompanyName>`
- **Category.Attribute**: Semantic subdivision that helps LLMs understand the context of anonymized data (as opposed to `[REDACTED]`)

### Open-Set Types

`--types` is not limited to predefined types — any natural language entity type can be specified (the model was trained on approximately 70,000 types). Parenthetical descriptions can be appended to type names to guide model focus, e.g., `"numeric values (transaction amounts)"`.

### Public/Private Distinction and Multilingual Support

- **Public/private distinction**: Achieved by specifying discriminative types — e.g., use `"personal location"` instead of `"location"` to only anonymize private addresses while preserving public place names (tested and reliably stable). ⚠️ Public/private distinction for person names (`"personal name"` vs `"person name"`) is **unstable** on the current 0.6B model and should not be relied upon
- **Multilingual**: Natively supports 8 languages: Chinese, English, French, German, Spanish, Portuguese, Japanese, and Korean. Cross-lingual text can be processed in mixed form

### Type Selection Guidelines

`--types` is flexibly determined by the Agent based on context:

- **User explicitly specifies** → follow user's request
- **Intent is clear, types are obvious** (e.g., "anonymize this contract" → names + organizations + amounts + addresses) → Agent decides autonomously
- **Intent is ambiguous or involves sensitive decisions** → first use `scan` to scan for as many entity types as possible, show discovered entities to the user for confirmation, then use `hide` to anonymize

## Prerequisites: llama-server (Auto-managed, Local-Only)

HaS depends on llama-server to l