SQLX

SQLX

[Listed on mcpservers.org]

Connect to MySQL, MariaDB, TiDB, GreatSQL, OceanBase, PostgreSQL, CockroachDB, YugabyteDB, openGauss, Oracle, SQL Server, ClickHouse, Trino, Presto, StarRocks, Apache Doris, TDengine, Dameng, KingbaseES, Apache Kylin, XuguDB, IBM Db2, IBM Informix, SUNDB, GBase 8s, Redis and MongoDB, or open a local SQLite, DuckDB or H2 file, from your terminal or from your agent. Connections are saved encrypted, one invocation runs one or more statements or commands, and the results come back complete and structured.

Quick start

# 1. Install the CLI (macOS, Linux and Windows x64; Node.js 22 or newer)npx -y @ottermind/sqlx@latestexport PATH="$HOME/.local/bin:$PATH"
# 2. Create a connection (an interactive terminal prompts for the username and password)sqlx datasource add --name dev --type postgresql --host db.example.com --port 5432 --database appsqlx datasource test --id dev
# 3. Run SQLsqlx sql execute --datasource dev --command "SELECT current_database()"

Using an agent? See Use with your agent: install one plugin or extension and the agent calls SQLX directly.

Install the CLI

Three channels, pick one: prebuilt packages from GitHub Releases, the npx installer, or the platform install scripts.

macOS and Linux

curl -fsSL https://raw.githubusercontent.com/OtterMind/sqlx/main/scripts/install.sh | shexport PATH="$HOME/.local/bin:$PATH"sqlx --versionsqlx init

The installer selects your platform, downloads the executable and verifies its SHA-256. It installs to ~/.local/bin; add that directory to your shell's persistent PATH. SQLX_INSTALL_DIR chooses another directory and SQLX_VERSION selects a release version.

Windows x64

Run in PowerShell:

$installer = Join-Path $env:TEMP 'sqlx-install.ps1'Invoke-WebRequest 'https://raw.githubusercontent.com/OtterMind/sqlx/main/scripts/install.ps1' -OutFile $installerpowershell -NoProfile -ExecutionPolicy Bypass -File $installer$env:Path = "$env:LOCALAPPDATA\Programs\SQLX;$env:Path"sqlx --versionsqlx init

Add %LOCALAPPDATA%\Programs\SQLX to your user PATH for future sessions. Neither installer overwrites an unrelated executable already named sqlx; use a different install directory in that case.

Prebuilt targets are macOS ARM64/x64, Linux ARM64/x64 and Windows x64; the Linux baseline is Ubuntu 24.04. Release users do not need Rust, Java or database drivers: database workers and a private JRE are downloaded only when needed.

Node.js (npx)

macOS, Linux and Windows x64 with Node.js 22 or newer:

npx -y @ottermind/sqlx@latestexport PATH="$HOME/.local/bin:$PATH"sqlx --version

The installer verifies the release manifest, SHA256SUMS and the downloaded archive before installing. It uses the same user-level location as the platform installers and installs the Skill into ./sqlx, which no agent watches: --target codex, --target claude, --target dsh, --target pi or --target <directory> place the Skill where it is actually discovered, and --cli-only installs the CLI alone. Any other first argument is run as a SQLX command instead of installing, so npx -y @ottermind/sqlx@latest mcp starts the MCP server. Node.js is needed only by the installer, never by the CLI.

Updates

The CLI updates its own executable:

sqlx update check       # report the latest stable version without installing itsqlx update install     # download, verify and replace the executablesqlx update status      # read local history without network access

sqlx update install --version <version> installs an exact stable version. Updates do not stop running SQL or the UI service, do not modify saved connections, and do not update Skills or plugins automatically; use sqlx skill update for managed Skills, and an installed UI plugin keeps the version you selected (the default interface moves with the CLI release).

Interactive use checks in the background at most once per day and only prints a notice on stderr. Piped and CI commands skip that check, and SQLX_NO_UPDATE_CHECK=1 disables it. Source builds and package-manager-owned paths keep their own installation method.

Use with your agent

All four are ready as soon as they are installed: the plugin or extension installs the sqlx CLI itself when it is missing, so you never install it separately. Codex and Claude additionally require the CLI to meet the plugin's minimum version and refuse to start an older incompatible MCP server silently.

Each of them also leaves the Skill installed, because the CLI alone gives an agent the tools but none of the engine knowledge. A plugin installs the Skill into that agent's own directory on first launch when it is not there yet, and says so on stderr while it does.

Codex

codex plugin marketplace add OtterMind/sqlx@pluginscodex plugin add sqlx@ottermind

The plugin starts sqlx mcp over MCP. Read-only operations run directly; the two execution tools (sqlx_sql_execute and sqlx_sql_view) are marked destructive and Codex asks for approval by default.

Claude Code

claude plugin marketplace add OtterMind/sqlx@pluginsclaude plugin install sqlx@ottermind

Headless runs need an explicit tool allowlist:

claude --allowedTools "mcp__plugin_sqlx_sqlx__*" -p "List my SQLX datasources"

DeepSeek Harness

dsh plugin --profile web add @ottermind/sqlx-dsh    # browser UIdsh plugin --profile tui add @ottermind/sqlx-dsh    # terminal UI

Plugins belong to a profile: install it into the profile you use, then restart dsh.

Pi

pi install npm:@ottermind/sqlx-pi

Any other MCP client

A client with no plugin system takes the server straight from the published package. It installs the sqlx CLI itself on first use, so this is the whole setup:

{  "mcpServers": {    "sqlx": {      "command": "npx",      "args": ["-y", "@ottermind/sqlx@latest", "mcp"]    }  }}

Point command at an installed binary instead to start without npx; a GUI application often has no ~/.local/bin on its PATH, so give the absolute path (~/.local/bin/sqlx on macOS and Linux, %LOCALAPPDATA%\Programs\SQLX\sqlx.exe on Windows) and keep "args": ["mcp"].

Such a client receives no Skill, so its agent runs the tools without the engine recipes. Install it where the client reads instructions from with sqlx skill install --path <directory>.

SQLX is also listed in the MCP Registry as io.github.OtterMind/sqlx, which points at a container image for clients that install from there: docker run -i --rm ghcr.io/ottermind/sqlx:latest. A container reaches only the SQL-execution path — the credential page, the result page and the local workbench are served on the container's own 127.0.0.1, and the data directory lives inside it, so mount a volume at /root/.sqlx to keep saved connections. Prefer the package above when the user should type credentials or read results.

Skill only

sqlx skill install --target codex     # Codex and dsh share ~/.agents/skillssqlx skill install --target claudesqlx skill install --target dshsqlx skill install --target pi
sqlx skill status                     # list managed installationssqlx skill update                     # update them, keeping local edits

For another agent, pass its skill directory with sqlx skill install --path <directory>.

Connections and SQL

Create a connection

sqlx datasource add --name dev --type postgresql --host db.example.com --port 5432 --database appsqlx datasource test --id devsqlx sql execute --datasource dev --command "SELECT current_database()" --command "SELECT 1"

--id and --datasource accept a stable datasource UUID or its unique name.

To let the user type the password in a local page, so it never enters the conversation or a tool response:

sqlx datasource add --ui --name dev --type postgresql --host db.example.com --port 5432 --database appsqlx datasource setup-status --request-id <request-id>

To edit an existing connection and keep its saved password unless you replace it: sqlx datasource update --id dev --ui.

Connections do not use TLS unless you ask for it: --tls defaults to disable. Add --tls verify-full to verify the database certificate; a database that serves no certificate your machine trusts then fails with invalid peer certificate: UnknownIssuer (MySQL), error performing TLS handshake (PostgreSQL) or a closed connection (Oracle), which is the usual case for a local container. Either choice is also available as an update:

sqlx datasource add --name dev --type mysql --host 127.0.0.1 --port 3306 --database app \  --username-env DB_USER --password-env DB_PASSWORD          # no TLS, the defaultsqlx datasource update --id dev --tls verify-full             # verify the server certificate

Credentials

Credentials come from named environment variables, hidden interactive prompts, or a connection JSON object on stdin. Never put a literal password in a command argument.

sqlx datasource add --name dev --connection-stdin

{  "database_type": "postgresql",  "host": "localhost",  "port": 5432,  "database": "app",  "service": "",  "username": "example_account",  "password": "replace_with_real_input",  "tls": "disable",  "properties": {}}

Datasource responses omit usernames, passwords and vendor properties.

Importing connections

A tool that already stores connections can hand all of them over in one versioned document, so a migration never needs a password in a command argument. examples/import-connections.json is a runnable starting point:

sqlx datasource import --stdin < connections.json

{  "version": 1,  "mode": "merge",  "datasources": [    {      "name": "dev",      "connection": {        "database_type": "postgresql",        "host": "localhost",        "port": 5432,        "database": "app",        "username": "example_account",        "password": "replace_with_real_input",        "tls": "disable"      }    }  ]}

Each connection is the same object --connection-stdin accepts. merge (the default) adds new names and updates existing ones in place while keeping their IDs, and it never deletes a stored connection. Every entry is validated on its own, so an unsupported engine or an incomplete connection is reported in skipped with its reason instead of failing the document:

{"success":true,"data":{"mode":"merge","added":1,"updated":0,"unchanged":0,"total":1,"datasources":[{"id":"...","name":"dev","connection":{"database_type":"postgresql"}}],"skipped":[{"name":"legacy","reason":"invalid_connection","detail":"..."}]}}

--dry-run validates and reports without storing anything, --strict refuses the whole document when any entry is invalid, and --file <path> reads the document from a file for callers that cannot pipe stdin. The store is written once, so an import is never half applied.

total counts every connection the import considered: the connections already stored plus the document's entries, including the ones it skipped. With one stored connection and two entries, one of them invalid, the response reports total: 3, added: 1 and one skipped entry.

Commands

Execution behavior

Each invocation owns one database connection. Repeated --command arguments execute in order, initially with autocommit, and stop at the first error; there is no implicit all-or-nothing transaction. Temporary tables and session variables do not survive another invocation. Do not submit client directives such as GO, DELIMITER or psql backslash commands.

--command is the current flag and --sql is still accepted as an alias.

Output is one JSON object with one item per executed statement: results[].stmt identifies the statement, cols lists the columns as [name, type] pairs (base64, boolean and json join as a third entry when values are not plain text), rows holds positional values and count holds the driver's row count. Duplicate labels remain distinct. Numbers are encoded as strings to preserve integer and decimal precision, and binary data uses Base64. A failed statement carries its own error with an outcome; statements that never ran are listed in skipped; the object ends with success, which is also the process exit status.

Only a preview travels through standard output. A result set with more rows than preview-rows (10 by default) or more than 16 KiB of values is written completely to results/<id>/<statement>-<result>.jsonl inside the result directory, and its item carries a file path plus the object's id. Read more rows from that file, or page them with sqlx results rows --id <id> --offset 100. The default result directory is a private directory inside the system temporary directory, so results disappear when the machine reboots; sqlx setting set results-dir ~/sqlx-results keeps them, and sqlx setting set results-retention-hours 0 stops the 24-hour cleanup.

A caller that cannot read files needs the whole result in one answer: --full prints every row and stores nothing, sqlx setting set result-mode full makes that the default for this machine, and the MCP tool sqlx_sql_execute takes "full": true for one call. A large result then fills the caller's own output budget, which is why the preview is the default.

--events prints the raw worker event stream instead (protocol_version, datasource_id, events, success) for scripts that parse it, and never stores anything. --preview <rows> overrides the preview size for one call.

Check the final success flag and the exit status, and never replay an uncertain write automatically.

Local pages

Let the user inspect results in a browser and enter the password there.

sqlx sql execute --datasource dev --command "SELECT id, name FROM users ORDER BY id" --view

SQLX executes once and returns a local URL. The page loads the results automatically, supports multiple result sets and pagination, and preserves exact values. Reloading, paging or reopening the page reads the cached result; Refresh on the page reruns the original SQL batch against the database, so any writes in that batch run again. A successful refresh replaces the displayed snapshot at the same URL; a failure keeps the previous result and does not roll back database changes. Results are retained locally for 24 hours.

sqlx ui opens the local workbench, sqlx ui status and sqlx ui stop inspect or stop the service, and --no-open returns a link without launching a browser. The workbench also renders the charts and dashboards created with sqlx analytics (see Local analytics); that command needs no running service. Pages are reachable only on the machine running SQLX, load an HttpOnly browser session automatically, and the service stays up until sqlx ui stop.

Typing the password into the page keeps credentials out of the conversation, but it does not isolate them from an agent that can read files or control the browser as the same operating-system user. See the local UI design for the interface and storage boundaries.

Choose your UI

sqlx ui plugin install --url <plugin-zip-url> --sha256 <published-sha256>sqlx ui plugin listsqlx ui plugin use <plugin-id>          # default restores the default interfacesqlx ui plugin remove <plugin-id> --version <version>

Installing does not activate a plugin; after selecting it, reload an open page or run sqlx ui. CLI updates keep the version of a plugin you installed and move the default interface to the version of the release, and switching versions needs sqlx ui stop first. Plugins run locally and can access entered credentials and displayed data, so install interfaces from authors you trust: a checksum proves the downloaded bytes are intact, not that the author is trustworthy.

To build your own interface, see the UI plugin guide, the typed browser SDK and the independent terminal UI example. Users need no Node.js runtime.

Data and downloads

User data lives in ~/.sqlx/; use --data-dir or SQLX_DATA_DIR for another location. Settings are stored in ~/.sqlx/settings.json and managed with sqlx setting list|get|set|unset; SQLX_PREVIEW_ROWS, SQLX_RESULTS_DIR, SQLX_RESULTS_RETENTION_HOURS and SQLX_RESULT_MODE override the file for one environment, and a command line flag overrides both. Stored results live in the result directory described above, keep 24 hours by default and stay under 1 GiB in total; older results are removed before the next command runs, and page results from --view are managed by the local service. Saved connections use AES-256-GCM with an independently generated local key: back up the key together with the encrypted data, because losing the key prevents decryption. Device identity is generated locally and this version uploads no device information.

The main executable contains no database drivers; each database's worker is downloaded on first use. MySQL, MariaDB, TiDB, GreatSQL, OceanBase, StarRocks and Apache Doris share the MySQL worker, Redis, MongoDB, SQLite and DuckDB run in their own native workers, PostgreSQL, CockroachDB and YugabyteDB share the PostgreSQL worker, and Oracle, SQL Server, ClickHouse, Trino, Presto, TDengine, openGauss, Dameng, KingbaseES, H2, Hive, Apache Kylin, XuguDB, IBM Db2, IBM Informix, SUNDB and GBase 8s use the JDBC worker (the database table lists which worker serves which database). Downloaded resources come from the fixed release manifest of the running CLI version and are verified before use; --manifest <https-url> selects another manifest or a local test server.

Downloads happen on first use and are cached afterwards. Each one prints Downloading … with speed and estimated time, and a final Downloaded … in 12.3s (390 KB/s) line on stderr; the progress line is refreshed only when stderr is a terminal, so piped JSON stays clean. An interrupted transfer is retried up to three times, and rerunning a failed command reuses every component that is already installed. To avoid waiting inside the first query or page:

sqlx prefetch mysql ui      # MySQL worker and the local browser UIsqlx prefetch all           # adds the PostgreSQL, CockroachDB, YugabyteDB, openGauss, MariaDB, TiDB, GreatSQL, OceanBase, StarRocks, Doris, Oracle, SQL Server, ClickHouse, Trino, Presto, Hive, Kylin, XuguDB, TDengine, Dameng, KingbaseES, Redis, MongoDB, SQLite, DuckDB and H2 components, the JDBC runtime and the JRE

The database references explain each SQL operation's purpose, parameters, result and official documentation link.

Troubleshooting

Build from source

Source development requires Git, Rust 1.95, Node.js 22 and the platform's native build tools; Node.js only builds UI plugin assets. For Oracle or SQL Server development, also install a Java 17 JDK and Maven.

On macOS or Linux:

git clone https://github.com/OtterMind/sqlx.gitcd sqlxnpm --prefix ui cinpm --prefix ui run buildcargo build --workspace --release --lockedexport PATH="$PWD/target/release:$PATH"export SQLX_WORKER_DIR="$PWD/target/release"sqlx --versionsqlx initsqlx ui plugin install --path ui/distsqlx ui plugin use default

On Windows PowerShell:

git clone https://github.com/OtterMind/sqlx.gitSet-Location sqlxnpm --prefix ui cinpm --prefix ui run buildcargo build --workspace --release --locked$env:Path = "$PWD\target\release;$env:Path"$env:SQLX_WORKER_DIR = "$PWD\target\release"sqlx --versionsqlx initsqlx ui plugin install --path ui/distsqlx ui plugin use default

Keep the checkout at that location, or copy the CLI and both native workers into a dedicated directory and update PATH and SQLX_WORKER_DIR accordingly. This source build makes MySQL and PostgreSQL usable without a published worker manifest.

For Oracle and SQL Server, build the JDBC worker and place its driver JARs alongside those workers. On macOS or Linux:

mvn -B -f java/jdbc/pom.xml packagecp java/jdbc/target/sqlx-jdbc-0.1.22.jar target/release/sqlx-jdbc.jarcurl -fL https://repo.maven.apache.org/maven2/com/oracle/database/jdbc/ojdbc11/23.6.0.24.10/ojdbc11-23.6.0.24.10.jar -o target/release/ojdbc.jarcurl -fL https://repo.maven.apache.org/maven2/com/microsoft/sqlserver/mssql-jdbc/12.10.1.jre11/mssql-jdbc-12.10.1.jre11.jar -o target/release/mssql-jdbc.jar

On Windows PowerShell:

mvn -B -f java/jdbc/pom.xml packageCopy-Item java/jdbc/target/sqlx-jdbc-0.1.22.jar target/release/sqlx-jdbc.jarInvoke-WebRequest 'https://repo.maven.apache.org/maven2/com/oracle/database/jdbc/ojdbc11/23.6.0.24.10/ojdbc11-23.6.0.24.10.jar' -OutFile target/release/ojdbc.jarInvoke-WebRequest 'https://repo.maven.apache.org/maven2/com/microsoft/sqlserver/mssql-jdbc/12.10.1.jre11/mssql-jdbc-12.10.1.jre11.jar' -OutFile target/release/mssql-jdbc.jar

Java 17 must be on PATH, or SQLX_JAVA_BIN can point to its executable. These manual dependencies are only needed for source development; a binary release downloads its private Java runtime and drivers automatically.

Development and validation

npm --prefix ui cinpm --prefix ui run buildcargo build --workspace --lockedcargo test --workspace --lockedcargo clippy --workspace --all-targets --locked -- -D warningsmvn -B -f java/jdbc/pom.xml verifypython3 tests/distribution.pypython3 tests/ui_lifecycle.pypython3 tests/ui_distribution.pypython3 tests/ui_plugins.pypython3 tests/updates.pydocker compose -f tests/compose.yaml up -d --wait mysql postgrespython3 tests/integration.pypython3 tests/ui_api.pydocker compose -f tests/compose.yaml down -v# one group at a time; StarRocks and Doris each need a frontend and a backenddocker compose -f tests/compose.yaml up -d --wait tidb yugabytedbpython3 tests/databases.py tidb yugabytedbdocker compose -f tests/compose.yaml down -vdocker compose -f tests/compose.yaml up -d --wait starrockspython3 tests/databases.py starrocksdocker compose -f tests/compose.yaml down -vdocker compose -f tests/compose.yaml up -d --wait dorispython3 tests/databases.py dorisdocker compose -f tests/compose.yaml down -vdocker compose -f tests/compose.yaml up -d --wait greatsql tdenginepython3 tests/databases.py greatsql tdenginedocker compose -f tests/compose.yaml down -v# Dameng and KingbaseES have no public image; point the fixtures at a local instance# and export SQLX_TEST_DAMENG_PASSWORD or SQLX_TEST_KINGBASE_PASSWORD when they differpython3 tests/databases.py dameng kingbasedocker compose -f tests/compose.yaml up -d --wait presto hivebash scripts/jdbc-fixture.sh prestobash scripts/jdbc-fixture.sh hivepython3 tests/databases.py presto hivedocker compose -f tests/compose.yaml down -v# Kylin, XuguDB, Db2 and Informix take turns because each needs several gigabytesdocker compose -f tests/compose.yaml up -d --wait kylinbash scripts/jdbc-fixture.sh kylinpython3 tests/databases.py kylindocker compose -f tests/compose.yaml down -v# GBase 8s takes the vendor driver and its own instance; SUNDB needs a licensed installation# and XuguDB a trial image whose password is known. tests/databases.py lists every variable.export SQLX_TEST_GBASE8S_DRIVER=~/gbasedbt-jdbc.jarpython3 tests/databases.py gbase8s

For local native workers, set SQLX_WORKER_DIR to the absolute target/debug directory. For JDBC development, that directory also contains sqlx-jdbc.jar and ojdbc.jar or mssql-jdbc.jar; SQLX_JAVA_BIN can select Java 17 explicitly. These overrides are for development, not prerequisites for release users. The fixture scripts use dedicated test containers and test-only credentials.

Drivers you provide

Most JDBC drivers ship with the release and are downloaded on first use. Vendors that do not allow their driver to be redistributed are not packaged: IBM Db2, IBM Informix, SUNDB and GBase 8s connect with a driver you install once from the vendor.

sqlx driver add --type db2 --jar ~/Downloads/jcc-12.1.0.0.jarsqlx driver listsqlx driver remove --type db2

sqlx driver add copies the jar into <data-dir>/drivers/<engine>/, checks that it really carries the engine's driver class, and every later command loads it from there. A jar you provide also wins over a released component, which is how a newer vendor driver is used before the release catches up. Running a statement for one of these engines without a driver explains the exact command to run:

sqlx sql execute --datasource <id> --command "SELECT 1 FROM SYSIBM.SYSDUMMY1"# SQLX does not redistribute the db2 driver; run `sqlx driver add --type db2 --jar <path>` with the vendor driver jar first

scripts/release-driver-check.sh <engine> <jar> <data-dir> <connection-json> <statement> runs the same path a release user takes — install the jar, then execute with no development overrides, so the CLI resolves the JRE and the JDBC runner from the published manifest. It needs a release whose JDBC runner already knows the engine, so run it against a published version rather than from a feature branch.

Where the drivers come from:

Releases and documentation

• Packages and checksums: GitHub Releases
• Release notes: docs/release-notes.md
• Design and implementation boundaries: docs/design.md
• Update behavior and recovery: docs/updates.md
• Local interface and plugins: docs/local-ui.md, docs/ui-plugins.md
• Agent integrations: integrations/README.md, Skill

Not included: SQL-file input, persistent sessions, configurable transaction and error modes, result-file export, unattended update installation, and telemetry.

License

This public source repository retains the modified Chat2DB license for adapted code. It is not the unmodified Apache 2.0 license. See LICENSE and NOTICE; third-party dependencies retain their own licenses.