A locally-mirrored research surface over four bulk academic archives of the Sumerian textual record — the dictionary, the annotated corpus, the literary canon, and the artifact catalogue — indexed for sub-millisecond query and exposed to language models via the Model Context Protocol.
The University of Pennsylvania's electronic Pennsylvania Sumerian Dictionary, second edition (ePSD2), is the standard modern lexical resource for Sumerian — published in 2017 by an international team led by Dr. Steve Tinney and integrated with the Open Richly Annotated Cuneiform Corpus (Oracc), a federated archive of roughly 138,000 transliterated cuneiform texts from museum collections in London, Berlin, Philadelphia, Istanbul, and Baghdad. Alongside Oracc, the Oxford Electronic Text Corpus of Sumerian Literature (ETCSL) contributes 394 hand-lemmatized literary compositions — hymns, myths, royal hymns, proverbs, the Sumerian King List, Inana's Descent, Gilgameš and the Underworld, the Šulgi praise poems — shipped with English translations alongside every Sumerian line. The Cuneiform Digital Library Initiative (CDLI) supplies catalogue metadata for 353,000 cuneiform-bearing artifacts: provenience, period, museum custody, dimensions, publication history, and links to the institutionally-hosted photographs and line drawings.
Each archive is distributed in a format suited to its own publication tradition — Oracc as 208 zipped JSON projects (~3.1 GB), ETCSL as TEI XML from the Oxford Text Archive, CDLI as a Git-LFS-hosted CSV — and excellent for almost nothing in particular until infrastructure is built on top of it. This project is that infrastructure. Streaming parsers ingest the source archives into denormalized SQLite indexes with FTS5 over Sumerian transliteration and English translation, bringing every operation — point lookup, attestation join, n-gram pattern query, line-with-lemma search — down to a local SQLite call typically under fifty milliseconds.
The surface this exposes is twofold: a Flask web application that recreates the canonical Oracc browsing experience offline with cuneiform-glyph rendering and attestation-line context, and a constellation of five Model Context Protocol servers organized along clean data-source boundaries (ePSD2, ETCSL, CDLI, OGSL sign list, and a teaching surface in the persona of Ummia 𒌝𒈪𒀀, the master teacher of the Sumerian scribal school). The MCP servers are designed around an attestation-first contract for language models: instead of letting LLMs confabulate plausible-looking morphology for a sparsely-trained, ergative-absolutive isolate, every tool returns real corpus forms ranked by frequency, with cited tablet sources, and every cited tablet links straight to its CDLI-hosted photograph. The agent's job is to choose; the corpus's job is to constrain.
This project is not a turnkey English-to-Sumerian translation system, and should not be used as one. Sumerian is an extinct, ergative-absolutive language isolate with a four-millennium attestation history spanning multiple dialects, registers, and orthographic conventions; competent translation requires a reader who can adjudicate among homographs, choose a historical period, weigh literary register against administrative register, and recognize when the corpus is silent on a construction. No language model — and no dictionary lookup chain, however well-grounded — can substitute for that judgment.
What this project does offer is a set of carefully indexed reference surfaces designed to aid the work of a human translator, student, or researcher who already brings (or is acquiring) that judgment. For the working scholar, the MCP servers function as a high-throughput lookup harness: attestation joins, n-gram pattern queries, period filters, and cross-references that would otherwise require hand-scraping Oracc's HTML pages. For students, the Ummia teaching server ships a structured five-lesson Sumerian 101 curriculum with a master-teacher persona, drill exercises, and seamless routing into the data servers for live vocabulary lookups and attested examples. For LLM agents, the attestation-first design constrains the model to real corpus forms — surfacing what scribes actually wrote so the human in the loop has authentic material to evaluate, rather than plausible-looking confabulation to cross-check.
The intended user is, in every case, a person doing the scholarship — the agent is a research instrument, not a replacement for the act of translation.
A local rendering of the electronic Pennsylvania Sumerian Dictionary (ePSD2) — provided strictly for 1:1 verification against the canonical site at oracc.museum.upenn.edu/epsd2. Not a replacement for the upstream dictionary; cite Oracc, not this mirror.
Five per-domain Model Context Protocol servers exposing the dictionary, the Oxford literary corpus, the CDLI artifact catalogue, and the Oracc sign list to AI agents.
Source code, build scripts, agent prompts, full setup guide.
Open source under MIT (code) with data layers separately
licensed — see LICENSE-DATA.md.
The Model Context Protocol is an open standard, originally introduced by Anthropic in November 2024, that lets language models call out to external tools, query databases, and read files through a uniform JSON-RPC interface — analogous to what a Language Server Protocol does for code editors, but for AI agents. An MCP server is the backend system the agent talks to; an MCP client is the agent runtime — Anthropic's Claude Desktop and Claude Code, OpenAI's ChatGPT (desktop and web, via its connectors / custom-MCP interface), Google's Gemini (via Gemini Code Assist and the Gemini CLI's MCP support), Cursor, Zed, or an LLM wrapper of your own. Eme-gir is that backend system for the Sumerian textual record — five independent MCP servers, one per data source, each surfacing its archive as a coherent set of tools the agent can invoke during a research or teaching session.
docker compose up.
Sumerian dictionary, attestation lookups, morphological decomposition, n-gram phrase patterns over 35.5M attestations.
Oxford literary corpus, 394 hand-lemmatized compositions with English translations — every result bilingual.
CDLI artifact catalogue lookup — provenience, period, museum custody, and links to institutionally-hosted photographs.
Cuneiform sign rendering via the Oracc Global Sign List. Reusable for Akkadian, Hittite, Hurrian, and Elamite — cuneiform is not Sumerian-specific.
Teaching surface in the persona of Ummia 𒌝𒈪𒀀, master teacher of the e₂-dub-ba; ships the Sumerian 101 curriculum and the dual-register grammar reference.
Drop this into ~/.config/claude/claude_desktop_config.json
(Claude Desktop), .mcp.json at your project root
(Claude Code), or the equivalent mcpServers map in your
MCP client of choice:
{ "mcpServers": { "eme-gir-epsd2": { "type": "http", "url": "https://epsd2.eme-gir.org/mcp" }, "eme-gir-etcsl": { "type": "http", "url": "https://etcsl.eme-gir.org/mcp" }, "eme-gir-cdli": { "type": "http", "url": "https://cdli.eme-gir.org/mcp" }, "eme-gir-ogsl": { "type": "http", "url": "https://ogsl.eme-gir.org/mcp" }, "eme-gir-ummia": { "type": "http", "url": "https://um-mi-a.eme-gir.org/mcp" } } }
Once connected, the agent should call start_here() on
each server at session start to load the per-server bootstrap
prompt (this is a no-op on subsequent calls). For an opinionated
end-to-end translation workflow, start with the Ummia teaching
server and call get_grammar_reference() to load the
dual academic + temple grammar into the agent's working memory.
Once your client is configured and the five servers are
reachable, the MCP handshake establishes a session — but the
agent is still naïve about how each server expects to be used.
Each server therefore ships a start_here() bootstrap
tool whose docstring is prefixed with the literal sentinel
⭐ CALL THIS FIRST, so a tools-list scan surfaces
it as the entry point. Calling it loads the per-server
operating instructions into the agent's context: how to compose
with sibling servers, which tools to reach for in what order,
what the response shapes mean, and (on Ummia) the teaching
persona itself. There are two recommended starting paths
depending on what you want to do.
Path A — Research & translation
For English ↔ Sumerian dictionary lookup, lexical research, lemma attestation, literary corpus queries, and artifact provenience research. Begin every session by issuing this single bootstrap prompt to your agent:
The agent will then have the operating context to answer queries by routing them to the appropriate server. Try:
lugal appears with the verb du₃."lugal-e e₂ mu-na-du₃ and parse the grammatical role of each token."Path B — Guided instruction
For learning Sumerian from the ground up — phonology, the verbal chain, the case system, modal prefixes, subordination — taught one-on-one by the master scribe of the temple school in the persona of Ummia. Begin with:
The agent adopts the teaching persona, surfaces the five-lesson Sumerian 101 curriculum, and waits for you to name where you'd like to begin. Try:
Both bootstrap calls are idempotent —
repeating either prompt within a session is cheap and never
disruptive. After the initial handshake, simply request
lookups, translations, attestations, or lessons in natural
language; the agent's loaded context dictates which server's
tools to invoke for any given question.
💡 Re-issue periodically. Long sessions cause
agents to drift as the conversation history grows — research
workflows start ignoring attestation-first guidance, skipping
the sense_count/sense_pct ranking,
or omitting cited links; the Ummia teaching persona's voice
flattens, the scribal frame slips, drills get skipped. When
you notice that drift, paste either bootstrap prompt again
to refresh the per-server operating instructions — the
start_here() calls won't disrupt the in-flight
conversation, they just re-anchor the agent's working
context.
None of these datasets is unconditionally public-domain. All three
have meaningful reuse conditions surfaced in every tool response and
documented in
LICENSE-DATA.md.
Click any license name below to reveal the full verbatim citation
that every MCP response carries in its attribution field.
| Source | License | Notes |
|---|---|---|
| Oracc / ePSD2 / OGSL |
CC BY-SA 3.0ePSD2 / Oracc: electronic Pennsylvania Sumerian Dictionary, 2nd edition (oracc.museum.upenn.edu/epsd2), prepared by Dr. Steve Tinney and the Oracc team at the University of Pennsylvania. OGSL — the Oracc Global Sign List — is maintained alongside ePSD2 by the same team. Data licensed under Creative Commons Attribution-ShareAlike 3.0 Unported (CC BY-SA 3.0); see oracc.museum.upenn.edu/doc/about/licensing. Attribution is required; substantial reuses must propagate the ShareAlike license. |
Attribution required; ShareAlike propagates to substantial reuses |
| ETCSL (Oxford) |
Academic copyrightBlack, J.A., Cunningham, G., Robson, E., and Zólyomi, G., The Electronic Text Corpus of Sumerian Literature (etcsl.orinst.ox.ac.uk), Oxford 1998-2006. © The Authors; used per the project's academic citation convention. ETCSL is NOT released under any Creative Commons license; redistribution is governed by traditional academic copyright with a citation request that is treated as non-optional under the project's stated terms. |
NOT under any Creative Commons license; citation is non-optional |
| CDLI |
MixedCDLI: Cuneiform Digital Library Initiative (cdli.earth), hosted by the Max Planck Institute for the History of Science (Berlin) since 2022; originally UCLA. Catalogue text is freely reusable per CDLI's terms of use with citation to CDLI as the source. Imagery on cdli.earth is NOT openly licensed — non-commercial use only, with image copyright resting variously with CDLI, the photographer, and the holding museum. This server hosts no imagery; all image URLs link directly to cdli.earth, where CDLI handles licensing + access control. |
Catalogue text reusable-with-citation; imagery non-commercial only. We host no imagery — all links point to cdli.earth |
eme-gir is open source. The code
in this repository is released under the
MIT License
— fork it, modify it, run a private deployment of your own.
Written entirely in Python 3.12+, containerized via Docker,
runnable as a complete eight-service stack with a single
docker compose up. The public eme-gir.org
instance is hosted by the maintainer as a free service for
the Sumerological and digital-humanities communities;
contributions, bug reports, questions, and sponsorships are
all warmly welcomed via GitHub Issues.
Full source on GitHub. The code is MIT-licensed, so a
private deployment of your own is one git clone
away. The bundled data layer carries separate licenses
(Oracc CC BY-SA 3.0, ETCSL academic, CDLI mixed —
see Data sources above).
Python 3.12+ throughout. Flask + gunicorn for the web
surfaces. FastMCP (the official MCP SDK) for the server
layer. SQLite with FTS5 indexes for the data layer.
Docker (linux/arm64) for orchestration. No compiled
extensions beyond what requirements.txt
pulls from PyPI.
git clone the repo, then
docker compose up. The compose file brings
up the full eight-service stack — five MCP servers, two
Flask web apps, one init container — bound to localhost.
~10 min on first boot to download and index the Oracc
corpus; subsequent boots are sub-second.
This eme-gir.org instance runs on a small
AWS Docker Swarm (single-node, Graviton arm64). The
service is free to use while the current infrastructure
accommodates demand — see the sponsor note below if
traffic outgrows that.
🤝 Get involved. Pull requests, bug reports, feature requests, and questions are all warmly welcomed at github.com/jenova-marie/eme-gir/issues. Sumerologists' eyes on the data layer are especially valuable — if a translation, sense ranking, attestation, or period attribution doesn't pass scholarly muster, the issue tracker is the right place to flag it. Contributors should follow the existing project conventions documented in CLAUDE.md and the changelog discipline in CHANGELOG.md.
💝 Sponsorship. Hosting the public instance is a personal investment that's sustainable at current usage levels. If demand scales beyond the single-node swarm's capacity — or if a research institution would like to underwrite a dedicated build of the corpus — sponsorship keeps the public surface free for everyone else. Open an issue to start the conversation.
{d}inana za₃-mi₂