semlayer

semlayer

The semantic layer that infers itself — skip the quarter of hand-writing dbt YAML. Fair Source (FSL-1.1-ALv2); the format and its schemas are Apache-2.0.

Point it at your warehouse. It profiles every column, discovers the foreign keys nobody declared, decodes the status columns, finds the business rules hiding in your aggregate tables, and writes the whole thing down as a portable semantic layer in an openly licensed format — with confidence, provenance, and lifecycle on every single claim — ready for any AI agent to consume over MCP.

pip install "semlayer[warehouses]"

semlayer init snowflake # generates the minimal-grant setup script

snowflake reads the environment; snowflake: reads

~/.snowflake/connections.toml; snowflake://user@account/db?role=.. is a URL

semlayer infer snowflake -o layer.yaml \ --context ./docs/ --context ./etl-repo/CLAUDE.md # optional: your wikis/dictionaries as priors semlayer review layer.yaml # accept/reject what the engine inferred semlayer mcp layer.yaml # serve it to Claude, Cursor, or any MCP client semlayer lint layer.yaml query.sql # check any SQL (yours or an agent's) against the layer semlayer drift layer.yaml snowflake # catch schema changes (cron- and CI-friendly)

Or start from an agent instead of a terminal: add the MCP server to Claude Desktop, Claude Code or Cursor and ask it what it can see — it discovers the warehouses configured on your machine and says what to do next, with no layer built yet. See docs/mcp.md.

In VS Code and Cursor, the same layer underlines the fan-out joins and missing required filters as you type — docs/vscode.md.

From a checkout (to develop on semlayer itself): pip install -e ".[warehouses]".

Why

AI agents fail on real warehouses: frontier models solved just 21.3% of Spider 2.0's enterprise-warehouse tasks at publication (vs ~91% on the earlier academic Spider 1.0) — and even today's best agentic scaffolds only reach ~30%. The fix is a semantic layer — but every existing tool (dbt, LookML, Cube, Snowflake semantic views) makes humans write it by hand, and it goes stale the day someone runs an ALTER TABLE.

On our messy-warehouse benchmark (cryptic names, zero declared constraints, hidden business rules), an agent using the inferred layer answers 82% of business questions correctly vs 42% from the raw schema (+95% relative), with or without the layer's SQL linter — and the errors it fixes are the *silent* kind: raw-schema "total revenue" happily sums cancelled orders; a fan-out join quietly triples a total. Full benchmark, methodology changes included, and where we DON'T help →

What gets inferred

Everything lands with confidence (calibrated, with the misses published), provenance (which signals produced it), and lifecycle (inferred → reviewed → certified, plus deprecated/orphaned), governed by a normative consumer contract that makes silent misuse — summing across a fan-out, joining SCD2 at current-row, filtering on guessed decodes — *non-conforming*, not merely unwise.

Cost & privacy

• ~$0.70 per 100 tables end-to-end on the cheap model tier, with your own API key (measured cost model). LLM calls are escalation-only; ~80% of columns resolve from statistics alone.
• Several teams, one surface: each team infers its own schema into its own file; an analyst's hub lists them as members and holds the cross-document relationships, semlayer stitch infers those through the analyst's own connection, and every command takes the hub (docs/sets.md).
• Tiers: every table carries a tier (core, staging, raw, scratch) inferred from its name and ETL columns; raw and scratch tables are modeling: inventory_only -- listed with their shape, never joined, described, metricised or queued for review column by column. On a 2,000-table synthetic estate that is 62% of tables set aside, the review queue cut 4.4x and foreign-key precision 0.31 → 0.94 (measured).
• `--no-llm`: fully deterministic mode, zero API calls, still useful (0.780 typing / 0.834 role across 7 fixtures, against 0.809 / 0.838 with the model) — for orgs where LLM access needs procurement.
• `--no-sample-egress`: no value read from a warehouse cell is sent to the LLM, at no net accuracy cost across our 7 fixtures (0.809 typing / 0.842 role, versus 0.809 / 0.838 with samples). Per warehouse it trades rather than being uniformly free — up to a point better or worse. Re-measured 2026-10-03; see docs/cost-model.md.
• Two knobs for speed, output identical at any value: --llm-workers N (default 4) runs that many model calls at once (typing escalation, foreign-key validation, both Describe passes; bounded by your API rate limit, measured 7.2x at 8) and --query-workers N runs that many foreign-key queries at once *in your warehouse* (default 4 on a network warehouse, since it spends your compute: measured 3.3x at 4 and 6.7x at 8 on BigQuery; every core on an in-process DuckDB file). --no-llm builds still use --query-workers. (measured)
• Read-only, minimal-grant: semlayer init generates the grant script; the reader persona is proved unable to write — on PostgreSQL by a test that runs in ordinary CI against a real server (tests/test_postgres.py, no credentials needed), and on Snowflake by a live suite runnable against your own account (tests/test_snowflake_live.py).
• Model-response cache, on your disk and nowhere else: each model response is cached in semlayer/cassettes under your platform cache directory ($XDG_CACHE_HOME or ~/.cache, ~/Library/Caches on macOS, %LOCALAPPDATA% on Windows), so an interrupted whole-warehouse run resumes without paying for the same prompts twice. The files hold what the model wrote *about* your schema — descriptions, inferred types, and the enum decodes it returned, which do quote the coded values it was shown (X=Cancelled). No credentials, no result sets, and nothing is sent to us. Switch it off with --no-llm-cache or SEMLAYER_LLM_CACHE=off, move it with SEMLAYER_LLM_CACHE= , or keep cell values out of the prompts entirely with --no-sample-egress.
• Telemetry: anonymous usage statistics — command, duration, outcome, warehouse type, bucketed table/column counts, version, OS, whether an infer was started from the CLI or from an agent, and for semlayer mcp which MCP client connected and which of our tools it called (one of seven fixed words for the client, never its own name; tool names only, never their arguments). Never names, SQL, or data. See exactly what is sent with semlayer telemetry --show; opt out with semlayer telemetry --off, or per-shell with SEMLAYER_TELEMETRY=off. (what and why)

Honest scope (beta)

• Best on messy warehouses. On clean, well-named schemas our benchmark shows raw DDL is already sufficient — we publish that negative result rather than hide it. If your warehouse is tidy TPC-DS, you may not need us.
• ETL input: dbt (--manifest target/manifest.json) -- reads the model DAG to settle join targets statistics cannot. Catalog comments and --context docs are read as priors on every run.
• Verify it yourself, free: ANTHROPIC_API_KEY= pytest -q replays recorded model responses, so every accuracy floor and benchmark number above is checkable with no API key and no spend — including from a fork, where CI has no secrets (what those recordings are, and what is still untidy about them). The warehouses behind the numbers, with sources, pinned versions and licences, are listed in docs/warehouses.md.
• Scale: measured to 50,000 tables, as one file or as one layer per team stitched into a hub; the largest real warehouse run so far is about 2,000. Data VOLUME is not the limit -- 4TB and 45TB are the same architecture. The ladder, and how to make it fast.
• Warehouses: Snowflake, BigQuery, PostgreSQL, DuckDB (+ Iceberg on S3 via the DuckDB bridge). PostgreSQL also reaches Redshift and other Postgres-protocol engines, whose catalogs differ in places and are untested. Exporter: dbt (losses reported, never silent). LLM: Anthropic API (Bedrock/Vertex routing is next).
• Not included: hosted service, ontology enrichment (it failed our own ablation gate — receipts), LookML/RDF exporters.
• The eval harness ships in this repo — fixtures, gold layers, competency questions, benchmark runner. Run our numbers yourself: python fixtures/build.py && pytest tests/ -q.

Beta feedback

Running the beta? File the Beta feedback form. GitHub issues are public and permanent, so the form asks only for numbers you are happy to say out loud — don't paste schema names, role names or the *.report.json the CLI writes next to your output. That file stays on your machine; it records which tables and schemas you scoped the run to, so it is yours, not ours. If you want us to see it, email hello@semlayer.dev — same address for anything else sensitive.

Layout

spec/ format schema + consumer contract · src/semlayer/ the engine · fixtures/ 9 eval warehouses + golds + CQ suites · docs/ benchmark, cost model, spike reports · ARCHITECTURE.md how it works

Licence

Fair Source, not open source. The engine and CLI are FSL-1.1-ALv2: free for any internal use, including in production and CI, with no fee or key; the one prohibition is offering semlayer to third parties as a competing product or service. Every release converts to Apache-2.0 two years after it is published.

The format specification, its JSON Schemas and the exporters are Apache-2.0 — a semantic layer nobody el