Htr Transcription

HTR Transcription

Transcribe handwritten historical documents using the HTRflow MCP server. Returns an interactive viewer, per-line transcription JSON, and archival exports.

Tools

• htr_transcribe — Transcribe images and return result URLs

Workflow

1. Determine image source

• http/https URLs (IIIF links, public image URLs): Use directly — skip to step 2.
• Local files or attachments: Must be uploaded first. Use the /upload-files skill, then continue to step 2.

2. Transcribe

Call htr_transcribe once with ALL image URLs in a single call.

Batching rule: Never call htr_transcribe multiple times for separate images. Each call runs an expensive GPU pipeline — batch everything.

3. Present results

After transcription, present results as an inline artifact for the viewer and downloadable links for data exports.

4a. Inline viewer artifact

Download the viewer HTML, then inline all external dependencies (OpenSeadragon JS and images) so the artifact is fully self-contained (the artifact sandbox blocks external requests).

curl -sL "{viewer_url}" -o /home/claude/viewer.html

Then run this Python script to embed dependencies:

import re, base64, urllib.request

with open("/home/claude/viewer.html", "r") as f:
    html = f.read()

# Inline OpenSeadragon JS (CDN script -> inline script)
osd_match = re.search(r'<script src="(https://cdn[^"]+openseadragon[^"]+)">\s*</script>', html)
if osd_match:
    with urllib.request.urlopen(osd_match.group(1)) as resp:
        osd_js = resp.read().decode()
    html = html.replace(osd_match.group(0), f"<script>{osd_js}</script>")

# Embed all Gradio image URLs as base64 data URIs
for url in set(re.findall(
    r'https://riksarkivet-htr-demo\.hf\.space/gradio_api/file=[^\s"]+\.(?:jpg|png)', html
)):
    with urllib.request.urlopen(url) as resp:
        img_data = resp.read()
    ext = "jpeg" if url.endswith(".jpg") else "png"
    data_uri = f"data:image/{ext};base64,{base64.b64encode(img_data).decode()}"
    html = html.replace(url, data_uri)

with open("/mnt/user-data/outputs/viewer.html", "w") as f:
    f.write(html)

Then call present_files with /mnt/user-data/outputs/viewer.html to render the interactive viewer as an inline artifact.

4b. Export links

Provide the remaining URLs as clickable download links:

• Transcription data: [pages_url] (per-line JSON)
• Export: [export_url] (archival export)

Do NOT reproduce document text as plain text in your response — present the artifact and links instead.

Options

Language

Layout

Export format

Custom pipeline

custom_yaml accepts a raw HTRflow YAML config string. Overrides language and layout. Use only when user explicitly provides one.

Example — English modern handwriting with a custom TrOCR model:

steps:
- step: Segmentation
  settings:
    model: yolo
    model_settings:
      model: Riksarkivet/yolov9-lines-within-regions-1
- step: TextRecognition
  settings:
    model: TrOCR
    model_settings:
      model: microsoft/trocr-base-handwritten
    generation_settings:
       batch_size: 16
- step: OrderLines