Semantic Rails

Agents need more than a SQL string. They need a way to see what exists, understand the allowed next move, validate the draft, and only then run it. Semantic Rails packages that loop for local projects, HTTP clients, and MCP hosts.

Semantic Rails is an Apache-2.0 open-source, agent-first semantic layer. It ships a Python package, a CLI, an MCP server, and an HTTP API. DuckDB works out of the box for local development; Snowflake, Postgres, BigQuery, Databricks, Athena, ClickHouse, MotherDuck, and DuckLake are available through optional connectors.

Project documentation is available in docs/ and at semantic-rails.com/docs.

Quickstart

Start from any empty project folder. This path uses the published PyPI package; no repo checkout is required.

mkdir semantic-rails-demo
cd semantic-rails-demo
uv venv .venv
source .venv/bin/activate
uv pip install 'semantic-rails>=0.2.0'

semantic-rails setup --interactive
semantic-rails repl --path ./my_package

If you use another Python environment manager, install Semantic Rails into that environment and then run the same setup command:

mkdir semantic-rails-demo
cd semantic-rails-demo
python3 -m venv .venv
source .venv/bin/activate
python -m pip install 'semantic-rails>=0.2.0'
semantic-rails setup --interactive
semantic-rails repl --path ./my_package

These onboarding commands require Semantic Rails 0.2.0 or newer. The setup wizard can create a starter package, set a local profile, validate the package, and add Query MCP or Architect MCP to Claude Desktop or Codex. On POSIX systems it can also start a managed local MCP server; on Windows it skips that offer and uses client-launched stdio or a foreground HTTP terminal instead. Inside the REPL, type author to create or update models, dimensions, times, measures, metrics, and segments with guided choices, pre-write previews, similar-definition warnings, parse validation, and session-local undo. After the wizard sets a local profile, most commands can omit --path:

semantic-rails ask "total amount by event type" --run
semantic-rails mcp setup

For a non-interactive setup path:

semantic-rails init my_package --yes
semantic-rails project validate --path ./my_package
semantic-rails ask --path ./my_package "total amount by event type" --run
semantic-rails mcp setup --path ./my_package

init and the wizard both create a runnable package with package.yml, graph.yml, models/, metrics/, examples/, tests/, and starter CSV data. Commands can pass --path ./my_package, run from inside the package directory, or use the local profile created by the wizard.

To turn the starter into your project, edit the package files in place:

• package.yml: package id, namespace, warehouse, and connection settings.

• graph.yml: entities and relationships between modeled tables.

• models/: source relations, dimensions, time roles, and measures.

• metrics/: governed metrics built from those measures.

• examples/ and tests/: known-good queries for validation and CI.

Then rerun:

semantic-rails project validate --path ./my_package
semantic-rails ask --path ./my_package "your business question" --run

For a guided authoring walkthrough, see docs/PACKAGE_AUTHORING.md. For MCP-assisted authoring, see docs/ARCHITECT_MCP.md.

Related MCP server: DataHub MCP Server

Local MCP

For your own package, prefer an absolute path because MCP clients may start from a different working directory than your shell.

PACKAGE_PATH="$(pwd)/my_package"
semantic-rails mcp setup --path "$PACKAGE_PATH"

mcp setup checks the package, confirms the MCP tools load, and previews the Claude/Codex config it would write. From inside the package directory, or after semantic-rails profile init --package-path ./my_package, you can omit --path.

To install local MCP config:

semantic-rails mcp setup --path "$PACKAGE_PATH" --client both --mcp both --install --yes

On POSIX systems, use the managed background lifecycle for an HTTP smoke:

semantic-rails mcp status --path "$PACKAGE_PATH"
semantic-rails mcp start --path "$PACKAGE_PATH" --port 8091
curl -s http://127.0.0.1:8091/health
semantic-rails mcp stop --path "$PACKAGE_PATH"

The wizard uses the default managed-server name. If it already started the server, mcp status will show it and mcp start returns already_running. For a stale registration, run semantic-rails mcp stop --name default; the interactive wizard also offers to remove a dead registration and retry.

Managed mcp start/status/stop requires POSIX process identity. On Windows, install the generated client config so Claude/Codex launches stdio, or keep the foreground HTTP process open in its own terminal:

$PACKAGE_PATH = (Resolve-Path .\my_package).Path
semantic-rails mcp setup --path "$PACKAGE_PATH" --install --yes
semantic-rails mcp http --path "$PACKAGE_PATH" --host 127.0.0.1 --port 8091

--mcp query installs the runtime/query MCP. --mcp architect installs the package-authoring MCP. --mcp both installs both. Raw mcp stdio and mcp http commands are foreground server processes; most users should let their MCP host launch stdio from generated config. mcp client-config remains available when you want to preview or write client config directly.

See docs/MCP_INTERFACE.md and docs/AGENT_QUICKSTART.md for the full MCP contract.

Source Checkout

Use the GitHub source path when you are contributing to this repository or want the checked-in examples under examples/.

git clone https://github.com/semantic-rails/semantic-rails.git
cd semantic-rails
uv sync --group dev
uv run semantic-rails packages
uv run semantic-rails query --package jaffle_shop --query-json '@examples/jaffle_shop_revenue_by_store.json' --verbosity minimal --sql-profile off

Expected package output includes the bundled synthetic fixture:

jaffle_shop

Contributor release smoke for the bundled package:

uv run semantic-rails parse-config --path configs/semantic_rails/jaffle_shop
uv run semantic-rails validate-config --path configs/semantic_rails/jaffle_shop --quiet

Build and verify the exact wheel and sdist before publishing:

uv build --out-dir dist
uv pip install --python .venv/bin/python --reinstall dist/semantic_rails-0.2.0-py3-none-any.whl
uv run python scripts/verify_package_distribution.py --dist-dir dist --no-build

Package Naming

Semantic Rails is the public product name. The PyPI distribution is semantic-rails, the Python import package is semantic_rails, and the CLI is semantic-rails. The published distribution also includes mf2sr, the MetricFlow translator behind semantic-rails import.

What It Provides

• A governed Query IR that can be discovered, planned, validated, compiled, and executed as separate steps.

• MCP tools for the agent loop: capabilities, catalog, discover, inspect, build-options, valid-values, plan, validate, compile, execute, and segment helpers.

• Structured errors for invalid metrics, dimension mismatches, bad filters, and policy failures, with recovery hints where possible.

• Local DuckDB by default, with optional warehouse connectors for production runtimes.

The bundled jaffle_shop package is useful for smoke tests and examples. It is not the onboarding path for modeling your own data; use semantic-rails init for that.

Warehouses

DuckDB is included. Install optional connectors only when you need them:

pip install 'semantic-rails[postgres]'    # also: snowflake, bigquery, databricks, athena, clickhouse
pip install 'semantic-rails[all]'         # every connector

MotherDuck and DuckLake use the core duckdb dependency. Package connection settings should keep secrets in environment variables or files, not YAML literals. See docs/DEPLOYMENT.md and docs/ADDING_A_DIALECT.md.

Docs

• Getting started

• Package authoring

• MCP interface

• Agent quickstart

• Architect MCP

• Query API

• Capabilities

• Benchmark evidence

• Deployment

• OSS release checklist

• Contributing

• Comparative capability evidence

Project Status

Semantic Rails is beta software. The open-source runtime, CLI, MCP stdio server, HTTP API, and DuckDB path are the supported core.

Support, issue reporting, conduct, and security reporting are documented in SUPPORT.md, CONTRIBUTING.md, CODE_OF_CONDUCT.md, and SECURITY.md.

License

Apache 2.0; see LICENSE.