wrenai

Wren AI CLI and Python SDK — semantic SQL layer for 20+ data sources.

Translate natural SQL queries through an MDL (Modeling Definition Language) semantic layer and execute them against your database. Powered by Apache DataFusion.

Installation

pip install wrenai                 # Core (DuckDB included)
pip install 'wrenai[postgres]'     # PostgreSQL
pip install 'wrenai[mysql]'        # MySQL
pip install 'wrenai[bigquery]'     # BigQuery
pip install 'wrenai[snowflake]'    # Snowflake
pip install 'wrenai[clickhouse]'   # ClickHouse
pip install 'wrenai[trino]'        # Trino
pip install 'wrenai[mssql]'        # SQL Server
pip install 'wrenai[databricks]'   # Databricks
pip install 'wrenai[redshift]'     # Redshift
pip install 'wrenai[spark]'        # Spark
pip install 'wrenai[athena]'       # Athena
pip install 'wrenai[oracle]'       # Oracle
pip install 'wrenai[memory]'       # Schema & query memory (LanceDB)
pip install 'wrenai[memory-onnx]'  # Same, torch-free (onnxruntime instead of PyTorch)
pip install 'wrenai[ui]'           # Browser-based profile form (starlette + uvicorn)
pip install 'wrenai[main]'         # memory + interactive prompts + ui
pip install 'wrenai[all]'          # All connectors + main

Requires Python 3.11+.

Quick start

1. Initialize a project — scaffolds a YAML-based MDL project:

mkdir my-project && cd my-project
wren context init

This creates wren_project.yml, models/, and views/. Edit wren_project.yml to set your data_source and add models under models/:

# wren_project.yml
schema_version: 2
name: my_project
catalog: wren
schema: public
data_source: postgres

# models/orders/metadata.yml
name: orders
table_reference:
  schema: mydb
  table: orders
columns:
  - name: order_id
    type: integer
  - name: customer_id
    type: integer
  - name: total
    type: double
  - name: status
    type: varchar
primary_key: order_id

Already have an MDL JSON? Import it directly: wren context init --from-mdl path/to/mdl.json

2. Configure a connection profile:

# Browser form (recommended, requires wrenai[ui])
wren profile add my-db --ui

# Interactive terminal prompts
wren profile add my-db --interactive

# Import from an existing connection file
wren profile add my-db --from-file connection_info.json

3. Build the manifest:

wren context build

This compiles YAML files into target/mdl.json. The CLI auto-discovers this file when you run queries from within the project directory.

4. Run queries:

wren --sql 'SELECT order_id FROM "orders" LIMIT 10'

wren walks up from the current directory to find wren_project.yml and uses target/mdl.json. You can also pass --mdl path/to/mdl.json explicitly.

For the full CLI reference and per-datasource connection field reference, see docs/cli.md and docs/connections.md.

4a. (Optional) Aggregation queries with cubes — define cubes under cubes/, then query them with a structured input instead of writing GROUP BY SQL by hand:

wren cube list
wren cube describe revenue
wren cube query --cube revenue --measures total --time-dimension "order_date:month"

The translator produces DATE_TRUNC / GROUP BY / WHERE clauses for you and runs them through the same engine path as wren --sql. See the Cube guide for full YAML structure and the CLI reference for all flags.

5. (Optional) Configure security policy — create ~/.wren/config.json:

{
  "strict_mode": true,
  "denied_functions": ["pg_read_file", "dblink", "lo_import"]
}

6. (Optional) Index schema for semantic search (requires wrenai[memory], or wrenai[memory-onnx] for a torch-free install):

wren memory index                              # index MDL schema
wren memory fetch -q "customer order price"    # fetch relevant schema context
wren memory store --nl "top customers" --sql "SELECT ..."  # store NL→SQL pair
wren memory recall -q "best customers"         # retrieve similar past queries
wren memory watch                              # auto-reindex on schema/query changes

memory-onnx runs the same weights through onnxruntime instead of PyTorch, so it produces the same vectors and existing indexes stay valid -- it is a much smaller install, not a different index. When both extras are present, onnx is used; set WREN_EMBEDDING_BACKEND=sentence-transformers to override, and wren memory status reports which one is live.

7. (Optional) Build a shareable GenBI app — turn the context layer into a browser-side dashboard (powered by wren-core-wasm) and deploy it to Vercel or Cloudflare Pages. The CLI owns the build instruction + deterministic state; an agent authors the app from it:

wren genbi build sales --prompt "orders dashboard" --data-mode snapshot  # print build instruction
# agent authors apps/sales/ from the instruction (mdl.json + data/*.parquet)
wren genbi register sales --data-mode snapshot   # record the app
wren genbi verify sales                          # preflight (files, MDL, data, secret scan)
wren genbi open sales                            # local preview
wren genbi deploy sales --provider vercel        # ship a shareable URL (preview; --prod for production)

Tokens come from the env / .env (VERCEL_TOKEN / CLOUDFLARE_API_TOKEN), never CLI flags; Cloudflare needs wrangler installed. See the GenBI guide and the CLI reference.

8. (Optional) Serve an MCP server — expose the project's query, schema, and knowledge tools to Claude Desktop/Code, Cursor, or any MCP client. Runs in-process against the compiled MDL — no ibis-server, no separate service:

wren serve mcp                                # stdio (default) — client spawns this as a child process
wren serve mcp --transport http --port 8080   # local Streamable HTTP for other clients

Requires wren context build to have already run and the mcp extra: pip install 'wrenai[mcp]'. See the MCP guide and the CLI reference for the full tool/resource list and client wiring.

Connection profiles

Profiles let you store named connection configurations in ~/.wren/profiles.yml and switch between them easily — useful when working across multiple databases or environments.

# Add a profile (browser form, interactive prompts, or file import)
wren profile add prod --ui                        # opens http://localhost:<port>
wren profile add staging --interactive            # terminal prompts
wren profile add local --from-file conn.json      # import existing file

# List and switch profiles
wren profile list                                 # * marks the active profile
wren profile switch prod

# Inspect a profile (sensitive fields masked)
wren profile debug prod

# Remove a profile
wren profile rm old-profile --force

The --ui flag opens a browser-based form that auto-derives fields from each datasource's schema — including file upload for BigQuery credentials, variant selection for Databricks/Redshift, and sensible defaults for all 20+ supported sources. Requires pip install 'wrenai[ui]'.

Once a profile is active, wren uses it automatically:

wren profile switch prod
wren --sql 'SELECT COUNT(*) FROM "orders"'        # connects using prod profile

Python SDK

import base64, orjson
from wren import WrenEngine, DataSource

manifest = { ... }  # your MDL dict
manifest_str = base64.b64encode(orjson.dumps(manifest)).decode()

with WrenEngine(manifest_str, DataSource.mysql, {"host": "...", ...}) as engine:
    result = engine.query('SELECT * FROM "orders" LIMIT 10')
    print(result.to_pandas())

Development

Prerequisites: just and uv. (Rust + Cargo are only needed for the local-engine recipes below.)

Standard setup (no Rust toolchain)

just install        # uv sync — pulls the prebuilt wren-core-py wheel from PyPI
just lint           # Ruff format check + lint
just format         # Auto-fix

just install is a plain uv sync: it installs the locked prebuilt wren-core-py engine binding and the development tools from uv's default dev dependency group. No compilation required. This is enough for all Python-side development. Use just install-extra <extra> or just install-all for data-source extras.

Engine development (changing the Rust core)

Only needed when you modify ../wren-core-py (or ../wren-core) and want core/wren to run against your local build. Requires Rust + Cargo.

just install-local    # uv sync + build the local wheel + overlay it into .venv
just use-local-core   # rebuild + re-overlay after each subsequent Rust change

The run recipes (just test*, just lint, just dev) use uv run --no-sync, so they never revert a locally overlaid engine back to the lockfile version. If dependencies change, re-run an install recipe first.

Profile web tests (test_profile_web.py) require wrenai[ui]:

uv sync --extra ui
uv run --no-sync pytest tests/test_profile_web.py -v

Publishing

./scripts/publish.sh            # Build + publish to PyPI
./scripts/publish.sh --test     # Build + publish to TestPyPI
./scripts/publish.sh --build    # Build only

Package rename: wren-engine → wrenai

Starting with the 0.7.0 release, this PyPI distribution is renamed from wren-engine to wrenai to align with the Wren AI brand. The legacy wren-engine project on PyPI is frozen at 0.6.x and will not receive further updates.

What stays the same

• The Python import path: import wren (and submodules under wren.*)
• The wren CLI entrypoint and every subcommand (wren query, wren context, wren profile, wren memory, …)
• All extras (postgres, mysql, bigquery, …, memory, ui, main, all)
• Configuration files under ~/.wren/ (profiles, memory, config)

Only the name you type after pip install is different.

Migration

pip uninstall wren-engine
pip install wrenai                  # or: pip install "wrenai[<extras>]"
wren --version                      # should print: wrenai X.Y.Z

If your project pinned wren-engine in a requirements.txt, pyproject.toml, or lockfile, replace it with wrenai and re-lock.

License

Apache-2.0