TL;DR
auto-ontology-mcp (Apache-2.0, an early personal project, not affiliated with NVIDIA; surfaced by André Lindenberg) ports NVIDIA Auto Ontology's semantic layer into an MCP server. Its design choice is the interesting part: every language step (naming business terms, reviewing joins, extracting facts, writing SQL) is done by the model in your MCP client, and the server keeps only deterministic work: scanning, profiling, key inference, validation, storage and read-only execution in local SQLite.
It does two jobs:
Text-to-SQL that fails loudly. The model builds an ontology over your database; before a query runs, the server checks it for unknown joins, near-miss literals, percent-vs-fraction thresholds and one-to-many fan-out.
Agent memory as typed records. What the assistant reads (receipts, papers, notes) is stored as entities with typed columns, concepts with typed relations, indexed passages, and a source with the quote behind each fact. The thesis, in the post's words: money and counts "belong in typed records, not prose".
How it works
Catalog and keys. The server reads tables through SQLAlchemy, samples rows for top values, null rates, ranges and small enum sets, and infers undeclared foreign keys by name and value containment, including roles like billing_ vs delivery_. On the bundled demo it found 9 undeclared keys, auto-accepted 7, sent 1 self-reference for review and reported 1 it could not place.
Modeling. The model gets one table at a time with samples and Auto Ontology's modeling rules, proposes Terms and column mappings, and the server validates and stores them. Terms carry draft, certified or deprecated status; editing a certified Term returns it to draft.
Answering. check_answerable matches a question's entities to Terms and stored values and says when nothing covers a concept ("return on ad spend" with no ad-spend data). validate_sql then checks the query against the ontology and a dry run, and run_sql executes it read-only. Verified question and SQL pairs feed later answers and mined reusable measures.
Knowledge store. One writable SQLite file holds entities, concepts, passages and sources. Full originals stay in an Obsidian vault and passages link back to the heading they came from. A changed document replaces only its own passages; nothing is rebuilt. Concepts, relations and FTS5 indexes are ordinary tables, so one validated query can combine a text search, a recursive walk over relations and the records. Search is lexical (bm25, trigram) because there is no embedding model.
Safety. Connected databases are read-only in three layers: a SQL guard that allows one SELECT/WITH per call against the scanned catalog and blocks file and server functions, read-only connections, and rollback of every transaction. Queries time out at 30 seconds and return at most 500 rows.
Tradeoffs and failure modes
Quality is the client model's. Terms, joins and extraction are only as good as the model doing them, and different clients will build different ontologies from the same database.
Extraction errors become authoritative. A misread receipt is a typed row that sums cleanly. Provenance (source and quote per record) is what makes review possible, but nothing forces it.
Lexical search only. No embeddings means synonyms and paraphrase depend on the model adding them to Terms; the server instructions push it to.
Credentials in plain text. Datasource URLs, passwords included, sit unencrypted in a local SQLite file; the README says to use read-only accounts.
Scope cut from the original. No users, RBAC, access zones, PII detection or embedding search. This is a single-user tool, not a team semantic layer.
Early and single-maintainer. Version 0.5-era, one author; the benchmark numbers in the README are the author's own demo runs.
Implications
The broader pattern is worth copying even if you never install it: put a deterministic validator between the agent and any number it reports, and keep countable facts in typed storage with provenance rather than in prose memory.
Decide which agent memory is countable (money, dates, counts) and give it a schema
Store a source and quote with every extracted fact, and surface it with answers
Validate generated SQL for joins, literals and fan-out before running it
Connect databases with read-only accounts only
Certify important definitions and send edits back through review
Re-check extraction on a sample of records before trusting sums
Related entries
R010 PageIndex: also replaces similarity search with structure the model navigates; here the structure is a typed schema rather than a document tree.
R005 Second Brain OS: the LLM-wiki pattern this project offers to replace with typed tables and linked concepts.
R002 The 5 layers of agent memory: entities, concepts and passages map onto semantic and episodic memory, with the reconciliation step built in.
R020 The model proposes, the harness enforces:
validate_sqlis a concrete harness-side check on model output.R025 LongCapHarness: the same split, with rule-like steps in code and judgement in the model.
Sources
gogoymh/auto-ontology-mcp README, main branch, 2026-11-10
André Lindenberg LinkedIn post, 2026-11-10