ashishsinha1602/schemagate

Your text-to-SQL agent picks which tables to show the model before anyone checks what the caller is allowed to read. schemagate does the check first: it filters the schema by the caller's grants, so restricted tables are absent from the prompt rather than ranked low. Works with LangChain, MCP, or any SQL agent, on Postgres, Oracle, MySQL, SQL Server and SQLite.

If it keeps a table out of a prompt it should never have reached, a ⭐ helps other people building SQL agents find it.

With row-level security alone the failure is quiet: the model writes valid SQL against a table the caller cannot read, RLS strips every row, and the user is told "no records found" — indistinguishable from "this data does not exist."

Demo · Install · Benchmarks · Local models · What it costs · Coming from Vanna

Same question, two callers, no database and no key:

schemagate demo "salary by employee"                                     # hr_compensation absent
schemagate demo "salary by employee" --principal okta:hr --role payroll  # now it is first

Absent, not ranked low. A table the caller may not read never enters the prompt, so no rewording of the question reaches it and there is nothing to filter out of the answer afterwards.

Try it in the browser — no install, no database, no model call.

The selection is a prompt, so the rest follows:

pip install schemagate
schemagate demo "which customers owe us money" --answer --provider <provider> --model <model-id>

main.crm_customer   main.crm_contact   main.v_customer_balance   (+5)
8 of 42 objects  ·  ~383 prompt tokens instead of ~2,036

-- SQL written by <provider> / <model-id>, from 8 tables
SELECT c.id, p.display_name, v.account_number, v.invoiced, v.paid,
       (v.invoiced - v.paid) AS balance_due
FROM v_customer_balance v
JOIN crm_customer c ON c.id = v.id_customer
JOIN core_party  p ON p.id = c.id_party
WHERE v.invoiced > v.paid

id  display_name       account_number  invoiced  paid     balance_due
--  -----------------  --------------  --------  -------  -----------
1   Northwind Trading  ACC-1001        33960.0   22080.0  11880.0
2   Kellner GmbH       ACC-1002        8760.0    3000.0   5760.0

Rows, from a question, with no database to set up — that runs against a bundled 42-object schema. <provider> is the model service you choose: openai, anthropic, gemini, oci, local (no key at all), or auto (whichever key is set). Point it at your own database with --url:

schemagate select "which customers owe us money" \
  --url "postgresql+psycopg://user:pw@host/db" --answer --provider <provider> --model <model-id>

No key? Drop --provider and it prints a prompt to paste into any chat, then run the SQL it gives you back with --sql "SELECT ...".

More of the bundled schema, with the questions people actually type:

schemagate demo "which customers owe us money"
schemagate demo "late shipments by carrier" --prompt   # the DDL the model gets

Against your own database it's the same shape:

schemagate select "revenue by month" --url postgresql://localhost/app --principal okta:jdoe --role finance
schemagate studio --url postgresql://localhost/app        # the same thing, as a page

schemagate studio opens a local page where you type questions, switch the caller's roles, edit hints, and watch what reaches the prompt and what doesn't. The same page runs publicly at https://ashishsinha1602.github.io/schemagate/ on the six bundled schemas, in your browser, with no server behind it. The selector on that page is a JavaScript port of this library, and a test runs both against 1,789 cases and requires identical rankings.

If you're coming from Vanna (archived March 2026), docs/migrating-from-vanna.md is the short version: Vanna applied identity when the SQL ran; schemagate applies it before the model sees the schema. Your User maps to a Principal in one line.

Every text-to-SQL call pays for the schema in the prompt. Dump the whole thing and you pay for every table on every question; hand the model six tables and you pay for six. Measured on the test schemas, average over their golden questions, same built-in estimator as tests/bench.py:

The last row is the one that matters: the selection stays around six tables no matter how big the schema is, so the saving grows with the schema. Real databases are the last row, not the first.

Worked example, with a price you should replace with your own: a 260-object schema, 5,000 questions a day, an input price of $3 per million tokens. Full schema: 16,095 × 5,000 × 30 = 2.4 billion tokens a month, about $7,200. With schemagate: 444 × 5,000 × 30 = 67 million, about $200. The browser demo has these two numbers as editable fields under the stats, so you can put in your own volume and price and watch it recompute against whatever question you ask.

Two more things that cost nothing here and money elsewhere: the selector itself never calls a model (BM25 plus a hashed embedder, offline, milliseconds), and the optional descriptions can be written by any chat window you already pay for instead of an API key — see Without an API key.

Two things go wrong when you point an LLM at a database schema.

The first is cost. Most systems paste the whole schema into the prompt on every question. That's fine for twenty tables and ruinous for two thousand.

The second is worse, and it's the reason I wrote this. Schema selection happens before the query runs, so it happens before row-level security can do anything. If your selection step isn't identity-aware, the model gets handed a table the caller can't read. It writes perfectly good SQL. RLS or VPD filters every row out. The user sees "no records found" and believes it.

That's not an access-denied message. It's a wrong answer with a confident tone, and the user has no way to tell the difference. Filtering the catalog by identity first is the only way I know to avoid it.

from schemagate import Catalog, Principal

cat = Catalog().bootstrap("postgresql://localhost/app")   # or: from schemagate.demo_schema import demo_catalog; cat = demo_catalog()
cat.hint("invoice_draft", "pre-issue drafts only, not real revenue")
cat.restrict("hr_compensation", ["payroll"])

sel = cat.select("revenue by month", top_k=6,
                 principal=Principal("okta:jdoe", roles={"finance"}))

sel.prompt_fragment()   # compact DDL, ready for the system prompt
sel.object_list         # [{'owner': ..., 'name': ...}]
sel.explain()           # why each object was picked

hr_compensation is not in that result and its name does not appear anywhere in the prompt text.

pip install schemagate

That's the whole thing. One dependency (SQLAlchemy), no API key, no model download. The default embedder is a hashed n-gram vectoriser that runs offline and gives byte-identical results on every machine.

Extras, all optional:

pip install 'schemagate[postgres]'     'schemagate[oracle]'
pip install 'schemagate[mssql]'        'schemagate[mysql]'
pip install 'schemagate[anthropic]'    'schemagate[openai]'      'schemagate[gemini]'
pip install 'schemagate[huggingface]'

huggingface is the no-key, nothing-leaves-the-machine path, and it is the one extra that is heavy: about 2 GB of wheels plus a 3.1 GB model download the first time you use it. It is deliberately kept out of schemagate[all]. docs/local-models.md has the whole story — the downloads, the load you wait through once, what it is good at and where it is worse than a hosted model.

Every release is signed. The wheels carry PEP 740 attestations — a signature from GitHub naming the workflow, repository and commit that built that exact file. Nothing is uploaded by hand and there is no API token to steal. Check one yourself with gh attestation verify <wheel> --repo ashishsinha1602/schemagate.

pip install schemagate            # add an extra for your driver, below
schemagate                        # opens http://127.0.0.1:8770 on the 42-object demo

Every wheel on PyPI carries a signed provenance attestation naming the commit that built it: gh attestation verify <wheel> --repo ashishsinha1602/schemagate.

Or without installing anything, with every driver already in the image:

docker run -p 8770:8770 -e SCHEMAGATE_DATABASE_URL=postgresql://…   ghcr.io/ashishsinha1602/schemagate

Leave the URL off and it opens on a 42-object sample schema with data in it, so there is something to ask questions of before you point it at your own.

Then, in the page:

1. Connect. Paste a URL — postgres://…, postgresql://…, mysql://…, oracle://… and a JDBC string all work, as does the wallet form for an Autonomous Database. Tick Save this connection and give it a name and the next start reconnects on its own.
2. Catalogue. Settings → Model → pick a provider, paste a key, Save model (it is saved, so a restart does not ask again). Then Catalogue this database in the rail. One sentence per object, cached to disk, so a second run costs nothing.
3. Ask. Type a question in your own words. You get the objects that answer it, the DDL a model would receive, the SQL, and the rows.

Drivers come as extras — schemagate[postgres], [oracle], [mysql], [mssql], or schemagate[all] for the lot:

pip install 'schemagate[postgres]'

Two levers, both per database and both applied on every reconnect:

• Hints (rail → Hints): one object, in your words. "MyConvo campaigns: personal-inbox sends from a user's own mailbox."
• Glossary (POST /api/glossary): one word, everywhere. A term here is fed to the cataloguing prompt, so every description uses your vocabulary, and expanded into questions that mention it.

A hint beats a generated description everywhere, and neither needs re-cataloguing.

Every subcommand, and what it is for. schemagate <command> --help prints the same thing.

schemagate demo      [question]          run against the bundled 42-object schema
schemagate select    --url URL [question]  select against your own database
schemagate studio    [--url URL]         the Studio page, served locally
schemagate describe  --url URL           write AI descriptions for your objects
schemagate certify   URL                 end-to-end check on a real engine

select is demo pointed at a real database; they take the same flags.

schemagate demo "who reports to whom"
schemagate select --url postgresql+psycopg://user:pw@host/db "unpaid invoices"
schemagate select --schema APP --url "oracle+oracledb://ADMIN:pw@/?dsn=(description=(address=(protocol=tcps)(port=1522)(host=adb.us-phoenix-1.oraclecloud.com))(connect_data=(service_name=abc123_mydb_low.adb.oraclecloud.com)))" "unpaid invoices"   # Autonomous Database

schemagate                              # the demo, or your remembered connection
schemagate studio                       # empty, connect from the page
schemagate studio --demo                # the bundled sample schema
schemagate studio --url postgresql://localhost/app
schemagate studio --remember            # save the connection, reconnect next time
schemagate studio --forget              # delete the saved connection and exit

In the page: Catalogue this database describes what has no description yet, Re-catalogue all rewrites every one, and Resync schema re-reflects the database while keeping the descriptions you already have.

# with a key
schemagate describe --url postgresql://localhost/app --provider anthropic --model claude-sonnet-5

# without one: write the prompt out, paste it into any chat, apply the reply
schemagate describe --url postgresql://localhost/app --out prompt.txt
schemagate describe --url postgresql://localhost/app --apply reply.json

schemagate certify "postgresql+psycopg://user:pw@host/db"

Reflects, selects, and reports what a real engine actually did — the check to run before trusting a new database or driver.

1. Reflect the schema through SQLAlchemy. No vendor SQL anywhere.
2. Index names, columns, comments, hints, and view definitions. That last one matters more than it sounds: a view exposes only its output columns, so v_stock_shortfall looks like it's about "shortfall" when the thing you'd search for, reorder_point, is buried in its SELECT.
3. Retrieve with reciprocal-rank fusion over BM25 and vector similarity. Neither alone is good enough. Vectors miss exact identifiers; BM25 misses "owe us money" → balance.
4. Walk foreign keys to pull in join tables the question never mentions. In my experience this is the single biggest cause of generated SQL that parses but won't run.
5. Apply the caller's identity at every step above.

On public benchmarks, so you can check them without trusting me: BENCHMARKS.md has schemagate on Spider and BIRD, with the scripts in benchmarks/ and the data downloaded from the original sources. The headline is the pooled setting — every Spider database merged into one 876-table catalog, no hint about which one to look in:

BIRD dev, 1,534 questions: 96.5% per-database at top_k=5, 91.1% pooled at top_k=10.

And Spider 2.0-lite, the benchmark built for real warehouses — 162 databases, 8,255 tables, a median of 16 per database and a maximum of 785: 79.8% at top_k=10 over all 247 usable questions, none excluded (84.5% on the 233 whose gold tables resolve), no pooling needed because the databases are already big. An earlier version of this page said 82.9%; that number was measured while Windows had silently made thousands of the schema files unreadable, and a second re-run was owed after the loader was found to be reading the benchmark's per-column descriptions as one table description. Both re-runs are done and BENCHMARKS.md keeps the corrections rather than deleting them.

End to end, on the metric those boards actually score: schemagate plus claude-opus-5 gets 68.0% execution accuracy on BIRD dev (102/150, seeded sample), with 97.3% of questions producing SQL that runs. The published GPT-4 baseline on BIRD dev is around 46%.

None of these is a leaderboard placing — that needs the held-out test set, and nothing here has been submitted.

Everything below is measured on schemas I invented, which is worth less and is why the public numbers come first.

Six test schemas ship with the library. Run python tests/bench.py and you get all of this printed back. TESTING.md is the full record of what was tested, what broke, and what was found to be the database rather than schemagate.

Every number below is printed by that run, and the run fails if any of them stops matching — bench.py reads this table back and compares.

recall@6 here is the share of gold tables retrieved in the top six, micro-averaged over questions. The literal column asks questions that reuse the schema's own vocabulary; the business words column asks for the same things the way a person does, with no vocabulary overlap. Both matter and they disagree, which is the point of showing both.

Measured with the hashed embedder — what pip install schemagate gives you, no extras. schemagate[huggingface] swaps in sentence-transformers and the business-word numbers move a long way: on the held-out paraphrase set tests/run_paraphrase_eval.py reports 58.6% overall hashed and 82.8% with MiniLM. That harness counts a question as hit if any gold table is retrieved, which is a looser predicate than this table's, so its figures are not comparable with these — it prints which embedder it used for the same reason.

Token counts come from an estimator built into the benchmark so the number is reproducible with no network and no extra install. pip install tiktoken and the same script switches to exact cl100k_base counts. The ratio holds either way.

Six schemas rather than one because a single schema whose questions happen to share vocabulary with its own table names will flatter any retriever. The second is a different domain entirely. The third is 260 objects of deliberate sabotage: an _archive and _stg copy of every table, the same table name in three schemas, an 8-deep foreign-key chain, a reference cycle, composite keys, a 320-column table, 100-character identifiers, and names in Spanish and Japanese. The fourth is a claims warehouse star schema built so that several tables are plausible for every question and one is right: the same fact at four grains, a slowly-changing member dimension with a history table, one date dimension joined five different ways, bridge tables, and fifteen _bkp, _old, _v2, _tmp and stg_ copies of the important ones. The fifth is a bank: a ledger at three grains, trades versus positions versus settlements, FX both as a daily table and an as-of view, lending, and the KYC and AML tables most callers must never see. The sixth is an IoT fleet: readings at raw, one-minute and hourly grains, six monthly partition tables, an alarm lifecycle spread across three tables. All six are invented. No real schema from anywhere is in this repo.

That 50% row is the honest one. Read it before you adopt this.

The default embedder matches subwords, not meaning. Ask it for "things we're running out of" and it will not find v_stock_shortfall, because those two strings have nothing in common. Ask it about stock_shortfall and it's excellent.

If your users type identifier-shaped questions, you're done, and you never need an API key. If they type like people, give the catalog descriptions. There are two ways, and neither is required.

With an API key in the environment, you get them without asking. Every path that answers a question — schemagate select, --answer, the MCP server, the Studio's connect — describes the catalogue first, caches the result per connection under ~/.schemagate/descriptions/, and re-describes an object only when its structure changes. A hint you wrote, or a database comment that says something, is never overwritten; a comment that only restates the object's name in the schema's own boilerplate is replaced, because it was diluting every word it contained. Measured on a 1,200-object schema, that is the difference between six and eight of eight complex questions producing SQL that runs. SCHEMAGATE_AUTO_DESCRIBE=0 turns it off; with no key present nothing is called and nothing changes.

Any chat window you already have — ChatGPT, Gemini, Copilot, a local model — can write the descriptions. schemagate gives you the prompt and takes the reply:

schemagate describe --url postgresql://localhost/app --out prompt.txt
# paste prompt.txt into a chat; save its JSON reply as reply.json
schemagate describe --url postgresql://localhost/app --apply reply.json --config catalog.json
schemagate select   --url postgresql://localhost/app "things we're running out of" --config catalog.json

The prompt is metadata only — names, types, comments, foreign keys, never rows — and one paste covers every undescribed object. The reply lands in the describe block of catalog.json, next to your restrict and hint blocks, and select, studio and the MCP server (SCHEMAGATE_CATALOG_CONFIG) all read it. From Python it's the same idea: cat.describe_prompt() and cat.describe({"v_stock_shortfall": "Items below their reorder level."}).

pip install 'schemagate[huggingface]'
schemagate describe --url postgresql://localhost/app                     --provider local --model Qwen/Qwen2.5-1.5B-Instruct                     --cache .schemagate-cache.json

or, in the Studio, Settings → Model → Local (transformers). No key field appears, because there is no key.

from schemagate.ai import SchemaDescriber, LocalProvider
cat.describe(SchemaDescriber(LocalProvider(), cache_path=".schemagate-cache.json"))

Writing one sentence per table is a small enough job that a 1.5B model does it acceptably. Writing multi-table SQL is not, and the Studio uses the same provider for both — so if you have a key, catalogue locally but answer with the key. Read docs/local-models.md before you turn it on: it covers the two downloads, the load you wait through once, the ~3 GB of RAM, the caching that makes the second run free, and why a weak model's bad description can no longer bury the object it describes.

from schemagate.ai import SchemaDescriber, AnthropicProvider

cat.describe(SchemaDescriber(AnthropicProvider(model="claude-sonnet-4-5"),
                             cache_path=".schemagate-cache.json"))

One sentence per table, written by the model, indexed like any other schema text. On the bundled schema that takes the business-words row from 50% to 100% with no change to the identifier-style questions.

Anthropic, OpenAI and Gemini are supported. Anything else goes through CallableProvider, which is also your escape hatch when a vendor changes their SDK and you don't want to wait for a release from me.

from schemagate.ai import (AnthropicProvider, OpenAIProvider, GeminiProvider,
                      CallableProvider, auto_provider, available_providers)

AnthropicProvider(model="claude-sonnet-4-5")                    # ANTHROPIC_API_KEY
OpenAIProvider(model="gpt-4.1-mini")                            # OPENAI_API_KEY
GeminiProvider(model="gemini-2.5-flash")                        # GEMINI_API_KEY
OpenAIProvider(model="…", base_url="http://localhost:11434/v1") # anything local
CallableProvider(lambda system, prompt: my_llm(system, prompt))

available_providers()      # ['AnthropicProvider'] — names, never key values
auto_provider(model="…")   # picks whichever key is set

model is required. I'm not shipping a default model ID, because model IDs change every few months and a hardcoded one eventually 404s for everybody who installed the version before the fix.

Three things worth knowing before you turn this on:

What leaves your network. Table names, column names, types, nullability, existing comments, foreign keys. Not one row of data — ObjectDoc has no field that could hold one, and there are tests asserting both halves of that. Nothing is sent unless you call describe().

What it costs. One short call per undescribed object, once. Objects that already have a database comment or a hint are skipped by default. Results cache by content, so re-running is free and only changed tables get re-described. Ask before you pay:

describer.estimate_calls(docs)   # calls describe() would actually bill for
describer.preview(doc)           # the exact text that would be sent

What happens when it fails. The object is skipped, cataloging continues, and describer.failures lists what was missed. Pass strict=True if you'd rather it raise. A hint() you wrote by hand always beats a generated description, so fixing a bad one costs nothing.

pip install schemagate uses the hashed n-gram vectoriser: offline, instant, byte-identical on every machine. Install schemagate[huggingface] and a sentence model is used automatically instead -- no flag, no benchmark, no decision for you. Measured across the six bundled schemas, 98 questions, no descriptions:

The gain is concentrated exactly where the hashed embedder is documented to be weak -- questions phrased the way people speak. On the commerce schema, which carries that set, it goes 15/18 to 18/18.

It is not the base default because that would trade one dependency for torch, and the guarantee that the same text gives the same vector everywhere. SCHEMAGATE_AUTO_EMBEDDER=0 keeps the hashed one if you need to reproduce an older index.

You can swap in a hosted embedder too, though on identifier-heavy schema text the offline ones are often just as good and cost nothing per query.

from schemagate.ai import APIEmbedder, OpenAIProvider

provider = OpenAIProvider(model="gpt-4.1-mini",
                          embed_model="text-embedding-3-small")
cat = Catalog(embedder=APIEmbedder(provider, dim=1536))

Some questions share nothing with any identifier, and no description or model reliably bridges them: "money we gave back to shoppers" is billing_credit_note. State it once, in the same catalog.json as your access rules:

{"terms": {"gave back": "billing_credit_note",
           "revenue": ["billing_invoice", "v_monthly_revenue"]}}

or cat.term("gave back", "billing_credit_note") in Python. A term matches as a phrase, plurals folded, and only the most specific term counts ("take things offline" beats "offline"). It is applied when the question is asked, so nothing is re-indexed, and it never widens access: a term for a table the caller may not read does nothing for that caller. A term naming an object that does not exist is an error, not a silent no-op.

What it is for, measured on the 58 business-language questions across the six bundled schemas (tests/run_terms_repair.py): every question that missed was fixed by adding one short term taken from it, and no other question broke -- 23 of 23 on the base embedder, 11 of 11 with MiniLM. What it is not: a glossary written in advance, from table names alone, barely moved the held-out questions (18/26 to 19/26 on the base embedder, 21/26 unchanged with MiniLM; tests/run_terms_eval.py), because people's words are hard to guess. Add terms for the questions you see missing.

A glossary term is the simplest concept. The full form says what a word means, precisely, so the model writes the right SQL rather than a plausible one:

{"ontology": {"concepts": {
  "revenue":  {"synonyms": ["turnover", "sales"],
               "maps": ["billing_invoice.total_net"],
               "filter": "billing_invoice.status = 'issued'",
               "definition": "Issued invoices only; drafts and void excluded.",
               "broader": ["money in"]},
  "money in": {"maps": ["v_monthly_revenue"]}}}}

A question that uses a concept's name or a synonym brings its objects in, and its meaning is written above the DDL:

-- What the business terms in this question mean:
-- revenue (also: turnover, sales): main.billing_invoice.total_net; only where billing_invoice.status = 'issued'. ...

Concepts one step away (broader, or narrower) add a weaker signal. A meaning line is left out for any caller who may not see everything it names -- the mapped table and column, and any table or column its rule or definition mentions -- so a concept never puts a withheld name back into the prompt.

Where a vocabulary comes from:

schemagate ontology import semantic_manifest.json --from dbt       --url ... --config catalog.json --save
schemagate ontology import model.yaml            --from snowflake --url ... --config catalog.json --save
schemagate ontology import glossary.csv          --from csv       --url ... --config catalog.json --save
schemagate ontology suggest history.jsonl        --url ... --config catalog.json         # review, then --save
schemagate ontology check                        --url ... --config catalog.json

suggest learns concepts from questions people asked and the SQL that answered them (--memory reads schemagate's own query memory). CSV headers from Collibra and Microsoft Purview exports work as they are.

What it does to accuracy, measured.

For scale on BIRD: pasting each question's own evidence into it (BIRD's usual setting) scores 61.2%, so the ontology delivers about 60% of that gain with no evidence typed per question; two identical runs without it differ by 18 answers. One result did not hold: concepts built from other questions' evidence gave 46.8% against 44.8% -- inside that noise -- because BIRD's evidence is mostly one-off formulas that rarely recur. An ontology helps with the definitions it actually holds.

Join-path completion through the schema graph was measured and not built: on Spider only 2 of 266 misses at top_k=5 are a missing bridge table, because foreign-key expansion already brings bridges in (benchmarks/join_paths.py).

Most people do not have a SQLAlchemy URL. They have a wallet zip, a JDBC string out of a config file, or a host and a port. schemagate.connect turns any of those into the two things SQLAlchemy needs:

from sqlalchemy import create_engine
from schemagate import Catalog
from schemagate.connect import resolve

url, connect_args = resolve("jdbc:oracle:thin:@//host:1521/ORCLPDB1")
cat = Catalog().bootstrap(create_engine(url, connect_args=connect_args))

connect_args is not optional. An Autonomous Database has no URL worth the name — the wallet directory, the wallet password and the TNS alias have nowhere to live in one — so the URL degenerates to oracle+oracledb://@ and the connection travels beside it:

url, connect_args = resolve({
    "kind": "wallet",
    "wallet": "~/Downloads/Wallet_mydb.zip",   # the zip as downloaded
    "alias": "mydb_high",
    "user": "ADMIN", "password": "...",
})

The zip is extracted next to itself, because the driver re-reads it on every reconnect — a temporary directory gives you a connection that works once.

Same thing from the Studio, with a dropdown instead of a dict:

pip install "schemagate[all]"          # every driver, the model SDKs, MCP
schemagate

On Oracle Cloud, add the OCI SDK so cataloguing can go through OCI Generative AI with no API key at all:

pip install "schemagate[all,oci]"

It is a separate word because it is a separate size: the OCI SDK is 488 MB and 17,505 modules, against 217 MB for everything else together.

Once connected, the Studio shows the command that reproduces it — the schemagate select ... --url ... line and the Python equivalent, with the password as $DB_PASSWORD rather than the real one. Try it in the page, then take the command.

Connecting works out of the box on localhost, where the only person who can reach the page is already sitting at a shell on that machine. Serve the Studio on any other address and it takes --allow-connect, because there it becomes a URL box anyone on the network can use to make your server connect to hosts only it can see. --no-connect turns it off anywhere. A failed connection reports the exception type and nothing else, because driver errors quote the URL they were given and a URL carries a password.

An object-level rule cannot express the common case: the table is the right answer and one column in it is not.

from schemagate import Catalog, Principal

cat = Catalog().bootstrap("postgresql+psycopg://user:pw@host/db")
cat.restrict_column("employee", "salary", ["payroll"])
cat.index()

analyst = Principal("okta:jdoe")
print(cat.select("who reports to whom", principal=analyst).prompt_fragment())

salary is absent from that fragment — not REDACTED, not renamed. The name is itself the disclosure: a model that knows the column exists can ask about it, join on it, or mention it in an explanation. Any -- FK line naming a withheld column is dropped too, since it would put the identifier straight back.

sel = cat.select("who reports to whom", principal=analyst)
sel.to_dict()
# {'question': 'who reports to whom',
#  'principal': 'okta:jdoe', 'roles': [], 'total_objects': 219,
#  'hits': [{'object': 'hr.employee', 'kind': 'TABLE', 'score': 0.031,
#            'reason': 'hybrid', 'columns_shown': 3, 'columns_withheld': 1}]}

The record an auditor asks for after the fact, and the one thing that cannot be reconstructed later — the catalog will have changed, roles will have changed, and the question is gone. It carries the count of withheld columns, never their names: a log that lists what it withheld has disclosed it to everyone who can read the log.

At forty tables a hand-written restrict map is fine. At four hundred it is a second copy of an ACL that already exists in the database, and two copies drift.

schemagate select "what do we pay our doctors" \
  --url "postgresql+psycopg://user:pw@host/db" --restrict-from-grants

from schemagate.grants import restrict_from_grants
report = restrict_from_grants(cat, engine, report=True)
# then the half grants cannot see: objects under a row-level policy are
# flagged in the prompt, probed per role on PostgreSQL, and the views that
# bypass the policy are named -- docs/row-level-security.md
from schemagate.rls import restrict_from_policies
print(restrict_from_policies(cat, engine, report=True))
print(report)
# postgresql: 629 object(s) seen, 218 restricted, 1 public, 0 unmatched, 1 role(s) expanded

PostgreSQL and Oracle. Nested roles are flattened transitively, so a user whose group maps to a role that inherits the granted one still reaches the object. An object with no grant row is left untouched and named in report.objects_unmatched — silence is not a denial, and restricting on absence would break a working catalog the first time a connection could not see everything.

It reads grants, so it does not see row-level policies. Measured on Oracle 26ai: a caller whose Virtual Private Database policy admits zero rows still holds SELECT in ALL_TAB_PRIVS, still appears in ALL_TABLES, and is still put in front of the model with every column. Nothing leaks -- the database enforces the policy -- but the prompt names a table that caller cannot get a row out of. docs/row-level-security.md has the measurement, what to do about it today, and the fix.

Grants answer which roles may see an object. The other half — which roles this caller holds — used to be whatever the caller said, which is fine for a desktop client on its own database and no check at all for a hosted server. A groups block in the catalog config reads it from where it is already kept, and the roles a client sends are then ignored:

{"groups": {"sources": [
   {"type": "entra", "tenant": "contoso.onmicrosoft.com",
    "client_id": "…", "client_secret": "${ENTRA_CLIENT_SECRET}"},
   {"type": "native"}],
  "map": {"Payroll Team": "payroll"}}}

Five sources: entra (Microsoft Entra ID through Graph, transitive group membership, ids and display names both), native (the database's own role graph — the same views restrict_from_grants reads, walked upward from the user, so a two-level GRANT chain resolves), sql (a membership table, one bound :subject), http (any endpoint returning groups as JSON), static (a mapping in the file). Each answers only for the subject namespaces it serves; results are a union; a map turns group ids into role names. Answers are cached for ttl seconds.

A source that cannot answer is an error to that caller, not an anonymous selection: "no groups" and "could not ask" are different answers and only one is safe to act on. Verified live on PostgreSQL 16, MySQL 8.4 and Oracle Autonomous Database 26ai: the resolver and the grants-restricted catalog agree with has_table_privilege and with an actual SELECT. docs/groups.md has the block, every source, and what was tested.

Reflection uses only SQLAlchemy's dialect-agnostic Inspector. There's no hand-written SQL in schemagate.introspect and a test fails the build if any appears, so in principle any dialect SQLAlchemy supports will work.

In principle isn't evidence, so there's a script:

python scripts/certify_dialect.py 'postgresql+psycopg://user:pw@host/db'
python scripts/certify_dialect.py 'oracle+oracledb://user:pw@host:1521/?service_name=FREEPDB1'
python scripts/certify_dialect.py 'mssql+pyodbc://user:pw@host/db?driver=ODBC+Driver+18+for+SQL+Server'
python scripts/certify_dialect.py 'mysql+pymysql://user:pw@host/db'

It creates three schemagate_cert_ tables, reflects them, runs selection and identity scoping end to end, drops them again, and exits non-zero if anything failed. Point it at a scratch schema.

Every row above has a real database behind it. The one thing still worth saying plainly: MariaDB is inferred from MySQL rather than run. Point the script at one and tell me what happens.

The same checks run under pytest if you export a URL, which is how CI certifies a dialect for good:

export SCHEMAGATE_POSTGRES_URL='postgresql+psycopg://…'
export SCHEMAGATE_ORACLE_URL='oracle+oracledb://…'
export SCHEMAGATE_MSSQL_URL='mssql+pyodbc://…'
export SCHEMAGATE_MYSQL_URL='mysql+pymysql://…'
pytest tests/test_dialects.py -v

The MCP server trusts the identity it is handed. principal and roles come from the client and are not authenticated -- there is no token and no session. Anyone who can reach the transport can claim a role and read what that role may read, and since run_query returns rows, that is data, not just schema. Run it over stdio (the caller is your own desktop client), or over HTTP behind something that authenticates the user and sets the principal for them. It is a scoping mechanism, not a lock. With a groups block in SCHEMAGATE_CATALOG_CONFIG the roles stop being the client's to claim — they come from the directory or the database and the ones in the request are ignored (docs/groups.md); the subject is still whatever the transport hands over.

If you already have an agent that writes SQL, the fastest way in is to let it call schemagate as a tool rather than wiring the library into your code.

MCP. Cursor, Windsurf, Zed, or anything else that speaks the Model Context Protocol:

One line, nothing to install first (needs uv); demo is a bundled sample schema, so this works before you point it at anything:

uvx --from "schemagate[mcp]" schemagate-mcp --url demo

Claude Code:

claude mcp add schemagate -- uvx --from "schemagate[mcp]" schemagate-mcp --url demo

Any other client (Claude Desktop, Cursor, Windsurf, Zed), with your database. Add your database's driver extra next to mcp: postgres, mysql, mssql or oracle (SQLite needs none):

{"mcpServers": {"schemagate": {
  "command": "uvx", "args": ["--from", "schemagate[mcp,postgres]", "schemagate-mcp"],
  "env": {"SCHEMAGATE_DATABASE_URL": "postgresql+psycopg://user:password@localhost/app"}}}}

Keep a URL with a password in env, as above, not in --url: command arguments are visible to every process on the machine. Without uv: pip install 'schemagate[mcp,postgres]', then run schemagate-mcp (or python -m schemagate.mcp_server, which still works).

Seven tools, all read-only: select_schema (the DDL for a question, scoped to the caller), list_objects (what this caller can see), describe_object (one object's full DDL), run_query (runs a single read, refused if it names anything the caller may not see), answer, refresh_catalog and health. They take principal and roles. If the client leaves them out, the caller is anonymous and sees only unrestricted objects. A restricted object and a missing one return the same error, so existence doesn't leak.

Every decision is recorded. Each call to those tools, and to run_query and answer, writes one line: when, which principal with which roles, what they asked, what they were shown, how many objects and columns were held back, what SQL ran and how many rows came back, and whether the call was refused and why. Never row data, never the names of what was withheld, never the database URL. It stays in memory (the last 500, counted in health) unless SCHEMAGATE_AUDIT_LOG=<path> — or =1 for ~/.schemagate/audit.jsonl — turns the file on; a tool whose pitch is that it stores nothing does not start writing files on its own. The log is for the operator, from the file; it is deliberately not a tool, because "recent decisions" handed to any client is every caller's questions handed to every other caller.

And the record can be proven, years later. Turn on signing and every record is chained to the one before it and signed with ML-DSA-65, the NIST post-quantum signature standard (FIPS 204): editing a record, deleting one, inserting one or reordering them is caught by schemagate audit verify, which names the line and the reason. Post-quantum because audit logs are kept for years, long enough that today's RSA and ECDSA signatures may be forgeable by then. It answers the auditor's question "can you prove the model was never shown the salary table for this user?" with a check anyone holding the public key can run.

pip install "schemagate[pq]"                       # PyCA cryptography >= 48 (OpenSSL)
schemagate audit keygen --out ./keys               # private key stays with the server
SCHEMAGATE_AUDIT_LOG=1 SCHEMAGATE_AUDIT_SIGNING_KEY=./keys/audit-signing-key.pem \
  python -m schemagate.mcp_server
schemagate audit verify ~/.schemagate/audit.jsonl --public-key ./keys/audit-signing-key.pub.pem
# OK  1204 signed records (ML-DSA-65), seq 0..1203, no edits, gaps or reordering
schemagate audit head ~/.schemagate/audit.jsonl     # keep this SEQ:HASH elsewhere

What it does not claim: someone who can write the file can still delete the newest records, because a shorter chain is still a valid chain. Keep the head somewhere else (a ticket, another system) and verify --head SEQ:HASH fails if the log no longer reaches it. Signing does not encrypt: the log is as readable as before, and still holds no rows and no names of what was withheld. It proves what schemagate decided, which is only as right as the roles it was given. One server process writes one log; a configured key that cannot be loaded stops the server at start-up rather than writing unsigned evidence.

It learns from SQL that ran. When answer produces a query that executes, the question and the query are remembered -- never the rows. The next similar question gets the tables that query read pinned into its selection, and the pair shown to the model as a worked example between the DDL and the question. Neither can widen what a caller sees: a pin goes through the same visibility gate as any pin, and an example is shown only when every table it names is visible to that caller. Every stored query is re-checked read-only on the way in and the way out. Memory-only unless SCHEMAGATE_MEMORY=<path> (or =1 for ~/.schemagate/memory/<db>.jsonl); with nothing remembered the prompt is byte-identical to the one before this existed. The CLI has --memory, the Studio uses it at connect. Similarity is the catalog's own embedder -- deterministic, offline, and a weak notion of "similar": it matches wording, not meaning, which is acceptable because the examples are advisory and the pins are gated. SCHEMAGATE_DATABASE_URL=demo serves the bundled schema. SCHEMAGATE_VALUES=1 is the server's --values: it reads the distinct values of short, non-personal columns at startup so the model writes 'CARD', not 'Card'. It reads rows, so it stays off unless set; SCHEMAGATE_VALUES_BUDGET caps it in seconds (default 30 -- raise it for a large schema behind a wallet).

To host it for a team rather than one desktop:

SCHEMAGATE_MCP_TRANSPORT=streamable-http SCHEMAGATE_MCP_PORT=8765 python -m schemagate.mcp_server

It's built not to die. The index lives in memory after startup, so the database going away does not take the server with it — select_schema keeps answering from the last good reflection, and refresh_catalog reports the failure instead of raising.

How old may the permissions be? The server reads them when it builds its catalogue: restrictions from the config file and, with SCHEMAGATE_RESTRICT_FROM_GRANTS=1, the database's own GRANTs. By default that catalogue is used until refresh_catalog is called, so a REVOKE in the database reaches the server only then. Set SCHEMAGATE_CATALOG_MAX_AGE=300 and a catalogue older than five minutes is re-read before it is used: a GRANT or a REVOKE takes effect within that window, not on the next call. If the re-read fails, the server does not keep answering from the old permissions: select_schema, list_objects, describe_object, run_query and answer refuse with "permissions are stale" until a re-read succeeds (retried at most every 5 seconds), and health shows catalog_stale and the reason. Caller roles from a groups resolver have their own bound, its ttl. Every tool catches everything and returns {"error": ...}; a bad request cannot end the session for other clients. health tells a load balancer what state it's in. A test throws 125 kinds of garbage at every tool and then checks the next good request still works, and another does the same through a real client over stdio. Works on MCP SDK 1.x and 2.x; the 2.0 rename broke a fresh install once and there's a shim and a test for it now.

LangChain. A proper BaseRetriever, so it composes:

pip install 'schemagate[langchain]'

from schemagate.integrations.langchain import SchemagateRetriever, prompt_fragment

retriever = SchemagateRetriever(catalog=cat, top_k=6,
                           principal=Principal("okta:jdoe", roles={"finance"}))
chain = retriever | RunnableLambda(prompt_fragment) | your_sql_prompt | llm

The principal is bound at construction on purpose. Build one retriever per caller; a chain can't forget to pass identity if the retriever already has it.

Certified live on Oracle AI Database 26ai. Two ways in, neither of which needs an API key — cataloguing runs on OCI Generative AI under your own OCI identity, so the prompts (schema metadata only, never rows) stay in your tenancy.

From Cloud Shell, about a minute, no VM:

pip install --user 'schemagate[oracle,oci]'
schemagate describe --url 'oracle+oracledb://@' --provider oci \
    --model google.gemini-2.5-pro --config catalog.json

Or one click, for an MCP endpoint that stays up for your team:

That opens Resource Manager in your own tenancy with the stack loaded — an Always-Free-eligible VM running the MCP server against an Autonomous Database it creates, or one you already have. Details and the Terraform: oci/.

Everything runs under your OCI identity and is billed to your tenancy. Nothing is installed in the database — no Select AI profile, no grants beyond SELECT.

pip install 'schemagate[oracle,oci,mcp]'

On Windows the OCI SDK has very long file paths: install into a short path (python -m venv C:\sg) or enable long paths, or pip fails part-way.

Sign-in is the usual OCI one: ~/.oci/config from an API key (Console › Profile › API keys), or a resource / instance principal when running inside OCI. Permission is one policy, for your group — no dynamic group, nothing on the database:

allow group <your-group> to use generative-ai-family in compartment <compartment>

Pick a model your region serves on demand. The catalogue in a region lists models that are dedicated-cluster only; calling one fails with 404 Entity with key <model> not found. Choose another (in us-phoenix-1, September 2026, google.gemini-2.5-flash was on demand; several listed Llama and Cohere models were not). The MCP server still starts when the model is wrong — it skips the descriptions — so check its log.

An MCP server on an Autonomous Database, with values and descriptions:

{"mcpServers": {"schemagate": {
  "command": "python", "args": ["-m", "schemagate.mcp_server"],
  "env": {
    "SCHEMAGATE_DATABASE_URL": "oracle+oracledb://@",
    "SCHEMAGATE_CONNECT_ARGS": "{\"config_dir\": \"C:/wallet\", \"wallet_location\": \"C:/wallet\", \"wallet_password\": \"...\", \"user\": \"APP\", \"password\": \"...\", \"dsn\": \"mydb_high\"}",
    "SCHEMAGATE_VALUES": "1",
    "SCHEMAGATE_VALUES_BUDGET": "120",
    "OCI_COMPARTMENT_ID": "ocid1.compartment.oc1..…",
    "OCI_REGION": "us-phoenix-1",
    "SCHEMAGATE_DESCRIBE_MODEL": "google.gemini-2.5-flash"
  }}}}

SCHEMAGATE_VALUES_BUDGET is seconds; the default 30 can cut value reading short on a large schema behind a wallet. Oracle-maintained schemas — including the SH and SSB samples every Autonomous Database ships — are left out, so the catalog is the schemas you own or were granted. From the CLI, --answer --provider oci --model <id> writes the SQL with the same model and runs it read-only over the same connection.

MemoryStore rebuilds on every process start. Fine for a few hundred objects, wrong for a long-lived service. OracleStore keeps vectors in Oracle 23ai's native VECTOR type so the nearest-neighbour search runs in the database:

from schemagate.stores.oracle import OracleStore

store = OracleStore(dsn="user/pw@host:1521/FREEPDB1", dim=512)
store.create_schema()                       # idempotent

cat = Catalog(store=store).bootstrap("oracle+oracledb://…")

Pass connection= instead of dsn= to reuse your app's pool. It won't close a connection it didn't open.

Scoping is a predicate inside the scored subquery, not a filter applied after the rows come back. A row the caller can't see is never ranked and never leaves the database.

Same caveat as above: 26 tests pin the SQL, the bind types and the scope predicate, and every statement is checked against an independent Oracle parser, but none of it has run against a live 23ai instance yet. To do that:

export SCHEMAGATE_ORACLE_DSN='user/password@host:1521/FREEPDB1'
pytest tests/test_store_conformance.py -v

Oracle Cloud's Always Free ATP is enough.

Archive and staging twins are handled, but know how. If your warehouse has orders, orders_bkp and stg_orders, the copies carry the same name words in a shorter document, and cosine similarity likes short documents. Left alone, a three-column _tmp copy beats the twenty-five-column table it was copied from, even with a hint on the real one — I watched it happen. So an object whose name is a real object's name plus _bkp, _old, _tmp, _v2, _archive and so on, or stg_/tmp_ in front, is ranked below the object it shadows. Only when that object exists: a lone pricing_v2 with no pricing is left alone. Only in the same schema. And never when you name the copy outright — asking for fact_claim_line_v2 gets you fact_claim_line_v2. The lists are DEFAULT_SHADOW_SUFFIXES and DEFAULT_SHADOW_PREFIXES; pass your own to Catalog(...), or empty tuples to switch it off. cat.shadows() shows what was detected.

Identifier length. PostgreSQL truncates names to 63 bytes at creation. That's the database doing it, not schemagate, and there's nothing to be done from this side.

Non-English schemas work, including Chinese, Japanese and Korean, and accents fold both ways so a search for facturacion finds facturación. But a question in English will not find a table named in Spanish. Nothing lexical can bridge that. Descriptions can.

top_k is not a hard cap. Foreign-key expansion runs after selection and adds join tables on top. That's deliberate — SQL that references a table you didn't include won't run — but size your prompt budget for it.

v1.0. The public API is stable.

The names this README documents -- Catalog and its methods, Principal, Selection, the schemagate CLI and its flags, the MCP tool names, the catalog config file and the LangChain retriever -- keep working without a breaking change until 2.0. Anything underscore-prefixed is internal and may move in any release.

Two things are deliberately not covered by that promise. The rows below marked unfinished, until they say done. And the ranking: retrieval is tuned release to release, so the set of tables a question returns can change between 1.x versions. What does not change is the part that matters -- a table the caller may not read is never in it.

import schemagate never imports any provider SDK, and there's a test asserting it.

Default embeddings are stable across processes, machines and Python versions, so cached or persisted vectors stay valid. That one is enforced by a test that runs the embedder in fresh subprocesses under different PYTHONHASHSEED values, because it was broken once and nothing else caught it.

Apache-2.0. Ashish Sinha.