Skip to content

Auto Ontology MCP Server Puts a Semantic Layer in Any MCP Client

A Python project on GitHub wraps NVIDIA Auto Ontology's semantic layer as an MCP server that hands every language step to the model already in your client.

Draftship Editorial3 min read

A semantic layer packaged for MCP clients

On 2026-11-14 the repository gogoymh/auto-ontology-mcp was on GitHub's trending list with 397 stars and 54 forks, gained in the eight days since it was created on 2026-11-06. It presents NVIDIA Auto Ontology's semantic layer as an ontology MCP server for other MCP clients, and its stated pitch is that no LLM API key is needed. The code is Python, licensed Apache-2.0.

Per the project's own description, the server has two uses. Point it at a database and the model you already run inside your client builds business terms, reviews joins and answers questions in SQL that the server checks against the ontology and the data. Or use it as a personal knowledge base: share receipts, statements, papers, book notes or screenshots and it stores facts as records, ideas as linked concepts, quotable passages in a full-text index, and where each came from. The clients the source names are Claude Code, Codex, OpenClaw, Hermes Agent, Cursor, VS Code, and the Claude and ChatGPT apps.

Which half the model does, which half the server does

Auto Ontology itself runs its own LLMs (NVIDIA NIM) behind a Next.js, FastAPI and Postgres stack. This project hands every language step, meaning reading, extracting, naming, judging and writing SQL, to the model inside your MCP client. That is why there is no model endpoint to host and no key to configure. The server keeps the mechanical work: schema extraction, profiling, key inference, validation, storage and read-only execution, in local SQLite files.

The target: SQL that runs and returns a wrong number

The repository's argument is that text-to-SQL fails silently. A query runs, a figure comes back, and nothing marks it wrong. The source lists the shapes it tries to catch: a model guessing whether billing_city_id or delivery_city_id means the customer's state; a join between customers.customer_id and orders.order_id made only because the integers overlap; a filter for 'pennsylvania' against a column storing 'Pennsylvania'; a threshold written as discount_pct > 20 on a column that runs from 0 to 0.3; prices summed after a join to order lines, inflating the total; and "return on ad spend" answered from whichever revenue column was nearby.

The tools are named in the source. validate_sql rejects a bad join and points at the one the ontology knows. check_answerable can return partially_answerable with a coverage figure and the terms it could not map. get_query_context and find_join_path supply join paths and role notes.

What the demo sets up

auto-ontology-mcp demo builds three demos around one fictional store: a synthetic retail database of 11 tables and 3,200 orders, three of the store's documents with originals in an Obsidian vault, and questions that need both. The README transcripts are abridged from runs of the test fixtures, and the replies shown are described as illustrative. demo --remove deletes what was added, --no-knowledge sets up the database only, and export_wiki writes the knowledge base out as an Obsidian vault with aliases, [[links]] and graph view.

Limits, and a likely reading

Two caveats come from the source itself. After a fresh scan the default city role is picked by an alphabetical fallback, so billing wins over delivery until someone calls set_relationship, and answers can rest on terms that are still drafts, which the demo output labels as unreviewed. The source does not say which database engines have been exercised beyond the SQLAlchemy reading, how much the local SQLite files can hold, or what happens with more than one user on the same machine. My reading, and this is inference rather than anything the source states: the project is likely most useful to someone who already keeps a database and a notes vault and wants one assistant over both, and less useful without an MCP client to run it in.

What it means for apps built with AI

Apps assembled with AI usually end up with a schema nobody has fully described, which is where quiet errors hide. The idea worth taking from this project is treating the schema and its joins as reviewable artefacts rather than context to paste in. On Draftship, apps come with a built-in database, and the page on managing your app data covers what your app holds. Draftship has no connection to this project and does not use it.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.