sscli
SQL Server CLI for AI coding agents.
One install. Your agents automatically know how to inspect SQL Server databases, run safe queries, and export results.
Why sscli?
| Token-efficient | Markdown output by default keeps agent context lean |
| Read-only default | Blocks INSERT, UPDATE, DELETE unless overridden (--allow-write) — no accidents |
| Single binary | Fast startup, no runtime dependencies |
| CLI over MCP | No "tool bloat" from verbose tool descriptions for tools that are rarely used |
| Progressive disclosure | Core commands visible, advanced disclosed when needed |
Why not sqlcmd
sqlcmd is a great general-purpose SQL Server client, especially for interactive sessions and ad-hoc work.
For tool-calling agents, sqlcmd tends to be a poor fit because it's optimized for humans, not for
structured, repeatable automation:
- Output is hard to consume:
sqlcmdoutput is human-oriented text; agents usually want stable markdown tables or a single JSON object they can reliably parse. - Schema discovery is manual: you end up writing catalog queries (
sys.tables,INFORMATION_SCHEMA, etc.) instead of calling purpose-built primitives likesscli tables,sscli describe, andsscli columns. - No safety guardrails:
sqlcmdwill happily run destructive statements if an agent makes a mistake.sscli sqlblocks writes by default and requires--allow-writefor mutations. - More setup friction:
sqlcmdis typically installed via Microsoft tooling and may require ODBC drivers depending on platform/CI image; sscli is a single binary with config + env var discovery built in. - No agent integration: sscli can install a reusable skill/extension so agents "know the tool" without you pasting usage docs into every prompt.
Keep sqlcmd for interactive SQL. Reach for sscli when you want safe-by-default queries, fast schema
inspection, and output formats that are easy for agents to use.
Quick Start (Agent Users)
1. Install sscli
# macOS/Linux
# Windows (PowerShell)
# or with cargo (any platform)
2. Teach your agents
Done. Your agents now know how to browse schemas, run safe queries, and export results.
What changes?
| Before | After |
|---|---|
| You paste schema context into prompts | Agent discovers schema on demand |
| Agent guesses at SQL Server commands | Agent knows sscli tables, sscli describe <Object>, etc. |
| Risk of accidental writes | Read-only by default, explicit --allow-write override |
| Verbose output bloats context | Token-efficient markdown output by default, --json if needed |
Manual Usage
For humans who want to use sscli directly.
1-minute setup (first run)
# Create a starter config in ./.sql-server/config.yaml (safe defaults)
# Set the password env var referenced by passwordEnv in your config. sscli also reads
# Sanity-check connectivity + server metadata
# See the effective settings + which config file was used
Common commands
Installation
Homebrew (macOS/Linux)
Scoop (Windows)
scoop bucket add jwcraig https://github.com/jwcraig/scoop-bucket
scoop install sscli
Quick install script
|
Cargo binstall (fast, no compilation)
From source
Prebuilt binaries
Download from GitHub Releases.
Development build
Updating
# Check if you're up to date (alias: `sscli upgrade`)
# Homebrew
# Cargo
Automatic update notifications (optional)
By default, sscli does not check for updates automatically.
To enable lightweight update notifications (stderr, TTY-only, cached), create:
~/.config/sscli/settings.json(Linux/XDG default)- macOS often uses
~/Library/Application Support/sscli/settings.jsonby default
Example settings.json:
Agent Integration
Supported agents
| Agent | Command | What it installs |
|---|---|---|
| Claude Code | sscli integrations skills add --global |
~/.claude/skills/sscli/SKILL.md |
| Codex | (same command) | ~/.codex/skills/sscli/SKILL.md |
| Gemini CLI | sscli integrations gemini add --global |
~/.gemini/extensions/sscli/ |
| Other agent harnesses | Via OpenSkills | Bridge to installed skills |
Per-project vs global
| Flag | Installs to | Use case |
|---|---|---|
--global |
~/.claude/skills/ |
Available in all projects |
| (none) | ./.claude/skills/ |
Project-specific override |
What the skill teaches agents
The installed skill file tells agents:
- When to use sscli (database inspection, schema discovery, safe queries)
- Available commands and their purpose
- Output preferences (markdown for context efficiency,
--jsonfor structured data) - Safety model (read-only default,
--allow-writefor mutations)
Configuration
sscli supports three ways to configure a connection (highest priority wins):
# 1) CLI flags (one-off / scripts)
# 2) Environment variables (CI-friendly)
# 3) Config file (recommended for repeated use)
&& &&
Creating a config file
Generate a commented template (writes ./.sql-server/config.yaml by default):
Or copy the example file in this repo:
Config discovery (where sscli looks)
--config <PATH>SQL_SERVER_CONFIG/SQLSERVER_CONFIG- Walk up from CWD looking for
.sql-server/config.{yaml,yml,json}or.sqlserver/config.{yaml,yml,json} - Global config:
$XDG_CONFIG_HOME/sql-server/config.{yaml,yml,json}(platform-dependent) - Environment variables
- Hardcoded defaults
Run sscli config to confirm which config file is being used and what values are in effect.
Example config.yaml
defaultProfile: default
profiles:
default:
server: localhost
port: 1433
database: master
user: sa
passwordEnv: SQL_PASSWORD
encrypt: true
trustCert: true
For a fully commented example (including settings.output.*, timeout, and defaultSchemas), see config.example.yaml.
Environment variables
Environment variables override values from the config file.
.env file support: sscli automatically loads a .env file from the current directory if present, reading any of the supported variables listed below. Use --env-file to load a different file (e.g., --env-file .env.dev). This is useful for local development without polluting your shell environment.
| Purpose | Environment variables (first match wins) |
|---|---|
| Config path | SQL_SERVER_CONFIG, SQLSERVER_CONFIG |
| Profile | SQL_SERVER_PROFILE, SQLSERVER_PROFILE |
| Connection URL | DATABASE_URL, DB_URL, SQLSERVER_URL |
| Server | SQL_SERVER, SQLSERVER_HOST, DB_HOST, MSSQL_HOST |
| Port | SQL_PORT, SQLSERVER_PORT, DB_PORT, MSSQL_PORT |
| Database | SQL_DATABASE, SQLSERVER_DB, DATABASE, DB_NAME, MSSQL_DATABASE |
| User | SQL_USER, SQLSERVER_USER, DB_USER, MSSQL_USER |
| Password | SQL_PASSWORD, SA_PASSWORD, MSSQL_SA_PASSWORD, SQLSERVER_PASSWORD, DB_PASSWORD, MSSQL_PASSWORD |
| Encrypt | SQL_ENCRYPT |
| Trust server certificate | SQL_TRUST_SERVER_CERTIFICATE |
| Connect timeout (ms) | SQL_CONNECT_TIMEOUT, DB_CONNECT_TIMEOUT |
sqlcmd compatibility: The following sqlcmd environment variables are also supported:
| Purpose | Variable |
|---|---|
| Server | SQLCMDSERVER |
| User | SQLCMDUSER |
| Password | SQLCMDPASSWORD |
| Database | SQLCMDDBNAME |
Commands
Core (shown in --help):
| Command | Purpose |
|---|---|
status |
Connectivity check |
databases |
List databases |
tables |
Browse tables and views (--describe for batch DDL) |
describe |
Any object: table, view, trigger, proc, function |
sql |
Execute read-only SQL |
table-data |
Sample rows from a table |
columns |
Find columns across tables |
Advanced (shown in help --all):
| Command | Purpose |
|---|---|
indexes |
Index details with usage stats |
foreign-keys |
Table relationships |
stored-procs |
List and execute read-only procedures |
sessions |
Active database sessions |
query-stats |
Top cached queries by resource usage |
backups |
Recent backup history |
integrations |
Install agent skills/extensions |
Note: sscli sessions filters by client host name using --client-host. --host is reserved as an alias for --server.
Output Formats
| Context | Default |
|---|---|
| Terminal (TTY) | Pretty tables |
| Piped / non-TTY | Markdown tables |
--json flag |
Stable JSON (v1 contract) |
--csv <file> |
CSV export |
JSON output emits exactly one object to stdout. Errors go to stderr.
Safety
sscli sql enforces read-only mode by default:
- Allowed: SELECT, WITH (CTEs), whitelisted stored procedures
- Blocked: INSERT, UPDATE, DELETE, DROP, ALTER, TRUNCATE, MERGE, etc.
Override with --allow-write when you intentionally need mutations.
JSON Contract (v1)
Each command returns a stable top-level object:
| Command | Shape |
|---|---|
status |
{ status, latencyMs, serverName, serverVersion, currentDatabase, timestamp, warnings } |
databases |
{ total, count, offset, limit, hasMore, nextOffset, databases: [...] } |
tables |
{ total, count, offset, limit, hasMore, nextOffset, tables: [...] } |
describe |
{ object: {schema, name, type}, columns, ddl?, indexes?, triggers?, foreignKeys?, constraints? } |
table-data |
{ table, columns, rows, total, offset, limit, hasMore, nextOffset } |
sql |
{ success, batches, resultSets, csvPaths? } |
Errors (stderr):
Testing
DB-backed integration tests (opt-in):
SSCLI_INTEGRATION_TESTS=1 SQL_SERVER_CONFIG=/path/to/config.yaml \
SQL_PASSWORD=...