sqlrooms CLI

Launch a local SQLRooms DuckDB project for adding data, authoring documents, and building Mosaic charts and dashboards.

Quick start

uvx sqlrooms ./sqlrooms.db

The sqlrooms distribution includes the CLI, UI, and reusable Python runtime. See migration instructions before upgrading an environment that contains the former sqlrooms-server distribution.

What happens:

• Runs one ASGI server: HTTP, DuckDB WebSockets at /ws/duckdb, the browser bridge at /ws/mcp-bridge, and optional HTTP MCP at /mcp share one port.
• Serves the SQLRooms document UI on http://localhost:3000, or the next free port, and opens your browser (disable with --no-open-browser).
• Drag-and-drop CSV, TSV, JSON, Parquet, and DuckDB files to load them into DuckDB; files are uploaded to a local sqlrooms_uploads folder and referenced by path.
• UI state is stored in the SQLRooms meta namespace (default __sqlrooms) of the selected DuckDB file.

Map basemaps

Maps use OpenFreeMap vector tiles with Positron for light mode and Dark for dark mode. No API key or registration is required.

New maps keep the light/dark style matching the app theme at creation, including after theme changes and workspace reloads. Change the saved style through Map settings → Basemap. Existing custom map styles are preserved.

CLI flags

• DB_PATH (positional): DuckDB project file to load/create (e.g. sqlrooms ./my.db). Required unless --db-path is provided.
• --version: Print the installed sqlrooms CLI version and exit.
• --db-path: DuckDB database to use as a flag alternative. Pass a filepath to persist, or :memory: for an explicit temporary in-memory session.
• --host / --port: HTTP host/port for the UI. The default bind address is 127.0.0.1. If --port is omitted, 3000 or the next free port is chosen automatically.
• --ws-port and --mcp-port: Removed. Use --port for all transports; old flags fail with migration guidance.
• --profile: Select a complete production capability profile: default, experimental, or document-charts-maps.
• --experimental: Compatibility alias for --profile experimental.
• --experimental-sync: Enable experimental sync (CRDT) over WebSocket (Loro). Requires the experimental profile.
• --ai-devtools: Enable the AI session devtools button in the UI, including production-built UI bundles. Can also be set with SQLROOMS_AI_DEVTOOLS=1.
• --debug: Enable verbose debug logging, including HTTP access logs and DuckDB query timing.
• --meta-db: Optional path to a dedicated DuckDB file for SQLRooms meta tables (UI state + CRDT snapshots). If omitted, meta tables are stored in the main DB.
• --meta-namespace (default __sqlrooms): Namespace for SQLRooms meta tables. If --meta-db is provided, used as ATTACH alias; otherwise used as a schema in the main DB.
• --no-open-browser: Skip automatically opening the browser tab.
• --ui: Optional path to a custom UI bundle directory (a Vite dist/). If omitted, uses the bundled default UI.
• --no-ui: Serve the same runtime/API without UI assets, equivalent to sqlrooms server --db-path ....
• --mcp: Enable /mcp on the shared listener, backed by the live browser room.
• --config: Path to a SQLRooms TOML config file. Defaults to ~/.config/sqlrooms/config.toml (%APPDATA%\sqlrooms\config.toml on Windows).
• --no-config: Disable config file loading.

Read-only artifact, document-block, and dashboard-panel image tools are always available in the CLI UI. Using their image results requires a vision-capable model and a provider that supports image tool results.

The server requires a loopback bind host. An explicitly configured development or external proxy must preserve the authenticated Host/Origin boundary.

The MCP listener uses the official stateless Streamable HTTP transport. The browser must remain open and initialized because the live room owns the tool catalog and execution state. Every MCP SQL query requires an allow-once dialog in that browser. This approval and the one-statement SELECT check are not a SQL sandbox; only approve SQL from a client and request you trust.

There is intentionally no sqlrooms add, sqlrooms import, or sqlrooms doctor command in the first public CLI. Drag-and-drop import is the supported first-launch path, and the release smoke checklist below covers the doctor-style checks for now.

Data persistence

Tables created in the selected DuckDB file (or attached meta DB if --meta-db is provided):

• __sqlrooms.ui_state (one row: key='default')
• __sqlrooms.sync_rooms (only used when --experimental --experimental-sync is enabled)

Uploads go to /api/upload. Runtime config for the UI is exposed at /api/config / /config.json.

Manual smoke test

Use this to prove the first-launch path:

uvx sqlrooms \
  --no-open-browser \
  ./smoke.duckdb

Then open the printed UI URL and verify:

• The app starts without a database connection error.
• Dragging a small CSV file into the data panel creates a table.
• The uploaded CSV lands next to smoke.duckdb under sqlrooms_uploads/.
• The data sidebar shows main.cars and does not show SQLRooms internal metadata.
• A document is created or selected automatically and contains a cars data-table explorer block.
• Users can create document and dashboard artifacts from the New menu without enabling --experimental.
• Map, notebook, canvas, app, HTML app, pivot, and SQL query surfaces stay hidden unless --experimental is provided.
• Restarting the same command with ./smoke.duckdb restores the imported table and persisted workspace state.

Config file

sqlrooms reads the app capability profile, AI provider settings, and connector settings from a TOML config file. AI settings changed in the CLI UI are saved back to this file automatically when config loading is enabled and the config file is writable:

• macOS / Linux: ~/.config/sqlrooms/config.toml
• Windows: %APPDATA%\sqlrooms\config.toml

Override with --config <path>, or disable with --no-config.

Example config file:

[app]
profile = "default"

[ai]
default_provider = "openai"
default_model = "gpt-5"

[[ai.providers]]
id = "openai"
base_url = "https://api.openai.com/v1"
api_key_env = "OPENAI_API_KEY"
models = ["gpt-5", "gpt-4.1"]

[[ai.providers]]
id = "anthropic"
base_url = "https://api.anthropic.com"
api_key_env = "ANTHROPIC_API_KEY"
models = ["claude-4-sonnet"]

[[ai.custom_models]]
model_name = "local-qwen"
base_url = "http://localhost:11434/v1"
api_key = "local-key"

[ai.model_parameters]
max_steps = 12
additional_instruction = "Prefer short answers."

[[db.connectors]]
id = "postgres-local"
engine = "postgres"
title = "Postgres Local"
host = "localhost"
port = "5432"
database = "postgres"
user = "postgres"
password = "postgres"

[[db.connectors]]
id = "snowflake-prod"
engine = "snowflake"
title = "Snowflake Prod"
account = "your-account"
user = "your-user"
password = "your-password"
warehouse = "your-warehouse"
database = "your-database"
schema = "your-schema"
role = "your-role"
authenticator = "externalbrowser"

[[db.connectors]]
id = "snowflake-dev"
engine = "snowflake"
title = "Snowflake Dev"
account = "your-dev-account"
user = "your-dev-user"
warehouse = "your-dev-warehouse"

Server-only mode (no UI)

Use the consolidated distribution and the same authenticated app:

uvx sqlrooms server --db-path ./sqlrooms.db --port 4000

Connect to ws://127.0.0.1:4000/ws/duckdb using the private native credential handoff described in AUTHENTICATION.md. No UI assets are required in this mode. MCP tools still require an owning browser; server-only mode does not make them headless. Use ./server to open a database literally named server, as with ./agent.

Backend connectors (DbSlice bridge)

Use these modes to run remote queries through backend connectors and materialize results into core DuckDB for downstream notebook cells.

Install optional connector dependencies first:

uv tool install "sqlrooms[connectors]"
# or install just one connector:
uv tool install "sqlrooms[postgres]"
uv tool install "sqlrooms[snowflake]"

Postgres

uvx sqlrooms \
  ./sqlrooms.db \
  --port 3000

Snowflake

uvx sqlrooms \
  ./sqlrooms.db \
  --port 3000

What this enables:

• sqlrooms exposes connector bridge endpoints under /api/db/*.
• Runtime connector metadata is exposed via /api/config, so frontend DbSlice auto-registers available backend connections.
• Notebook SQL cells can select Postgres/Snowflake connectors from the connector dropdown.
• Arrow payloads are materialized into DuckDB and can be queried downstream in the same session.

Notes:

• Configure connectors in sqlrooms.toml using [[db.connectors]] entries.
• Connector libraries are optional extras (postgres, snowflake, or connectors).

Interactive Claude Code workspace

# Claude must already be installed and authenticated (claude auth login).
# Options precede the positional database path, following the CLI's parser.
sqlrooms --claude --profile document-charts-maps ./my-project.duckdb

--claude enables MCP and external execution mode, opens the browser, waits up to 90 seconds for its authenticated workspace bridge, then runs a normal interactive Claude terminal session with inherited stdin/stdout. DuckDB opens an existing database or creates a missing file, just as in normal CLI startup; :memory: is also supported for a temporary workspace. A terminal is required. --no-open-browser is supported if you open the temporary single-use link printed on the interactive terminal; --no-ui is incompatible. No SQLRooms AI configuration is needed (--no-config is optional).

Claude retains its own authentication and model preferences. The launcher loads the bundled native plugin and session-only MCP configuration; it does not edit Claude's global configuration or select an evaluation model. Use /sqlrooms:sqlrooms to load the document/chart/map workflow. Only this session's SQLRooms MCP server is attached. The child receives a private credential-file path; a scoped header helper reads it to authenticate MCP. Tokens never appear in command arguments or shared agent configuration. See local authentication for browser launch tickets, native clients, expiry, and platform support.

The browser owns the workspace. Keep it open; manually edited content is visible through MCP. Verified reads of workspace tables run without a prompt. External or unverified SELECTs and database-writing commands require per-request browser approval. Use db.import-file to materialize local CSV/Parquet/JSON files and db.create-table-from-query for derived tables; both preserve existing tables unless replacement is explicit. The CLI does not expose room.add-url-data-source. See data import and approvals. Commands retain their existing validation. Disconnects fail pending operations; the launcher reports disconnect/reconnect without replaying edits. On Claude exit or cancellation, the launcher stops the HTTP/MCP listeners and reaps its Claude child. It does not close browser windows or terminate unrelated sessions. The DuckDB backend thread shares the launcher process lifetime, as in ordinary CLI launches. Forced termination of arbitrary Claude-spawned descendants is not claimed; Claude owns its native tool/subagent lifecycle.

For a browser managed by another client, use sqlrooms --execution-mode external --mcp ./existing.duckdb. Execution mode is independent of --profile. External mode composes no SQLRooms AI, AI-settings, or artifact/chat slice. Existing saved conversations and settings pass through workspace persistence unchanged until embedded mode is used again. This retains the existing persistence mechanism; it is not a new durability guarantee.

pnpm --filter sqlrooms-python build:ui prepares both the UI bundle and the plugin from the canonical CLI skill. Plugin generation runs after the cached UI build, so this step also supports CI and deployment paths that invoke uv build directly. pnpm --filter sqlrooms-python build prepares these assets, then packages and verifies them in the wheel. No marketplace installation or publication is needed. See the verification record for tested behavior and outstanding real-Claude authentication requirements.

All CLI listeners require authentication, including localhost callers. Public base URLs show a bootstrap recovery screen. See local authentication for the private native credential file and development proxy configuration.

Agent-managed workspaces

Run sqlrooms agent setup --client claude-desktop or --client claude-code to preview one-time integration. The stable sqlrooms agent connect MCP adapter can list saved projects, create named persistent workspaces, reuse live browsers, and reopen by saved workspace ID. sqlrooms agent status reports redacted diagnostics. Claude Code setup offers a recommended trust option for the known SQLRooms MCP tools, avoiding duplicate Claude prompts while retaining SQLRooms browser approvals. Use --no-trust-tools to opt out; existing user permissions and restrictions are preserved. See tool permissions.

See agent-managed workspaces for setup, profiles, explicit instance routing, recovery, lifecycle, and current verification limitations.

Reusable ASGI runtime

The core imports no CLI, UI assets, profiles or sync implementation unless enabled:

from pathlib import Path
from sqlrooms.server.access import LocalAccess
from sqlrooms.server.app import create_app, UVICORN_OPTIONS
from sqlrooms.server.runtime import DuckDBRuntime
from sqlrooms.server.security import TransportSecurity

access = LocalAccess()
runtime = DuckDBRuntime("analysis.duckdb", Path("./storage"), extensions=[])
security = TransportSecurity(access, {"http://127.0.0.1:3000"}, {"127.0.0.1:3000"})
app = create_app(runtime, security, title="My analysis app")
# uvicorn.run(app, host="127.0.0.1", port=3000, **UVICORN_OPTIONS)

create_app(..., configure=callable, lifespan=async_context_manager) lets the caller register application routes/assets after protocol routes and attach resources to the shared lifespan. sync_enabled=True opts into CRDT. The caller provides its own credential delivery; never embed the native token in browser source or URLs. DuckDBRuntime.start(), run_db_task(callable, query_id=...), cancel_query(id), ready, and close() own the database, per-operation cursors and executor. Cancellation is best effort and never promises to undo committed statements.

Run one Uvicorn worker per writable workspace. Limits: 128 MiB input frames, 128 MiB/64 outgoing messages per connection (including the in-flight frame), 32 pending operations/128 MiB retained input per connection, 64 database operations, and 64 connections including authentication handshakes. Overflow closes the affected slow client with 1013; sends time out after 15 seconds. WebSocket compression is explicitly disabled because large-frame compression blocks the shared event loop and delays control traffic. Ping interval and timeout are 20 seconds. Results are reauthorized before enqueue and send. Metadata and final checkpoint errors propagate instead of reporting a save. Loro stays installed to preserve the tested sync path; disabled sync initializes no Loro documents or background tasks. Pandas remains required for JSON encoding.

Alternate local applications

sqlrooms.agent.settings.ApplicationSettings is the explicit boundary for application identity, storage roots, managed launch policy and the browser tool contract. Pass one settings instance to Catalog, Manager (via its catalog), Runtime, Connector, and registry/development-launch helpers. Defaults preserve SQLRooms profiles, paths and commands. A fixed alternate application can provide an empty profile tuple, its own product/environment prefix and generated tool contract. No global environment rewriting is needed; identities and registry verification include the product as well as contract compatibility.

sqlrooms.server.bootstrap.register_bootstrap_routes installs the shared ticket exchange, renewal and native bootstrap routes. The composing application still must enforce the operation-specific TransportSecurity checks in middleware. create_app(..., distribution="sqlrooms") controls /version metadata; alternate applications supply their own installed distribution name. DuckDBRuntime(..., sync_storage=False) skips creation of CRDT storage tables; combine it with create_app(..., sync_enabled=False) to omit sync initialization and routes. Neither setting removes installed Python dependencies.

Roomie under python/roomie is a concrete alternate composition. It owns its application UI/routes and __roomie metadata, while reusing the server, access, MCP and managed-lifecycle implementations. The two applications are tested in one process to prevent identity or storage leaks.