300 tools is more than most clients will accept and far more than any one task needs. The split entry points exist precisely for that — transcription work is 47 tools, search is 16 — and choosing one is the difference between a working setup and a client that refuses to connect.
A wrapper over the Transkribus REST API for handwritten and printed text recognition. It covers collections and documents, page transcripts, recognition and layout-analysis runs, model training and management, search and keyword spotting, and the jobs, credits and uploads underneath all of it.
- 300 tools across 22 resource domains, published behind 8 entry points so you can load only the part your client has room for
- Collections and documents — create, list, search, duplicate and export collections; create documents from IIIF, METS, PDF or FTP; manage per-collection users, labels, tags and credits
- Pages and transcripts — read a page's plain text or a specific transcript version, post a new transcript, lock a page, move or replace pages
- Recognition — run ATR, OCR, CITlab HTR, PyLaia and TRHTR against a document, plus text-to-image alignment, and compute word error rate against ground truth
- Layout analysis — standard and advanced analysis, LA2 inference and table structure inference, with a cost estimate available first
- Training — CITlab HTR, LA2, PyLaia, P2PaLA and table-recognition training runs, with training and validation data and their statistics readable back
- Models — list, get, update, delete, and share a model into other collections
- Search — full text across transcriptions, keyword search, search and replace, tag search, and keyword-spotting queries with their hits
- Operations — jobs (create, list, restart, kill, undo, error detail), actions, users, credits and uploads
The npm package @lazyants/transkribus-mcp-server over stdio. Authentication is session-based: set TRANSKRIBUS_USER and TRANSKRIBUS_PASSWORD and the server logs in and manages the session, or set TRANSKRIBUS_SESSION_ID if you already hold one — session IDs expire, so credentials are the choice for anything long-running. Pick a split binary rather than the full server if your client caps tool counts: transkribus-mcp-collections is 132 tools, admin 62, transcription 47, users 29, models 26, jobs 18, search 16. Scope is the legacy TrpServer REST API; the newer Processing API v2 is deliberately out of scope. This is an unofficial server, not affiliated with READ-COOP SCE.
One command plus a key — npm install -g @lazyants/transkribus-mcp-server, then supply credentials