SQL Context Pack

SQL Context Pack turns a live database into AI-ready SQL context: the DDL and metadata of every TABLE, PROCEDURE and FUNCTION the owner's profile allows, plus a bounded masked sample of table data, organised into context folders — with anything it cannot confidently classify kept in unknowns/ rather than guessed.

Version 3.3.1 · Python >=3.11 · Output format 2

What it does

• Reads SQL Server, PostgreSQL, MySQL, MariaDB and Oracle, read-only.

• Exports DDL and metadata for tables, stored procedures and stored functions. Table data is only ever a bounded, masked sample — never every row.

• Writes a managed header into every SQL file carrying identity, context, description, tags, classification status and content hashes.

• Classifies files from an owner-registered folder, writing to a separate output directory by default.

• Maintains [agrimap_app].[DB_METADATA_CONTEXT] as a searchable context index, synchronised on every export so the next run starts better informed.

• Plans and applies procedure and function updates, one file or a whole folder, always through CREATE OR ALTER and always behind an approval.

• Runs the same Skill and MCP surface on Codex, Claude Code, Antigravity (agy) and Gemini CLI.

• Runs one service per machine. Every session connects through its own stdio bridge, not its own server.

Related MCP server: MCP Database Query Server

Boundaries worth knowing up front

• Credentials stay on the owner's side. The agent never receives them.

• The service binds loopback only.

• Both database write scopes are off by default, and are independent.

• Supervised database writes require a single-use, request-bound approval.

• Query Data is read-only SELECT, validated by parsing, with per-column masking.

• The context index and routine deployment are SQL Server features. The other four engines export full context and support Query Data, and say so plainly rather than failing oddly.

Repository and verification

Canonical repository: https://github.com/cognivex006/sql-context-pack. GitHub Actions are disabled and no Actions workflows are shipped. Maintainers run pwsh -NoProfile -File scripts/dev-check.ps1 and python scripts/validate_manifests.py locally before release. Version 3.3.1 is republished from this repository after migration.

Start here

1. Getting Started — from nothing to a first export.

2. Pick your host: Codex · Claude Code · Antigravity · Gemini CLI

3. Usage Examples — three worked flows, simple to advanced.

The full documentation map lists everything else.

Version 3.3.1

SQL Server fix: table catalogs no longer fail analysis on databases whose collation is not Latin1 (for example Thai_CI_AS). Read the release notes.

Version 3.3.0

Formatting fitted to each of the five dialects: sqlctx format takes its dialect from --profile (else tsql), and never applies rules that change meaning, such as rewriting NVL/IFNULL/ISNULL to COALESCE. Read the release notes.

Version 3.2.1

Oracle fix: table, procedure and function definitions are read for the requested object. Before, the schema and name reached DBMS_METADATA.GET_DDL swapped. Read the release notes.

Version 3.2.0

Security hardening: SQL Server connection values are escaped, Query Data proves a read-only login on every engine, locking reads are refused, format diagnostics are redacted but traceable, and Oracle profiles declare a connect mode. Read the release notes.

Version 3.1.0

SQL Server table exports are lossless and product-neutral: full types, identity, computed columns, defaults, collation, key order, FK actions, index INCLUDE/filter and column descriptions, in standard T-SQL and in the table_metadata YAML. Read the release notes.

Version 3.0.0

Sensitive data never reaches an AI model. Every value is classified and protected in the service before it is returned or exported: names, Thai IDs, phone, card and account numbers become marked fakes, emails and usernames aliases, address/birth date/location are generalized, and secrets are redacted. Protected columns cannot be used in filters or functions to infer values, and when real values are needed the AI hands the user a query to run themselves (sqlctx query --reveal, interactive terminal only).

It is a major version because Query Data output changes meaning and some queries 2.1.0 accepted are now rejected. Read the release and migration notes.

Version 2.1.0

2.1.0 is the first released version; the repository history begins there. It closes three Query Data read-path bypasses, makes the context index cover every discovered object on every export, adds the Antigravity host, and makes the shared service manageable from an installed package on every platform.

It is a major version because sqlctx_sync_context_index gained a mode and Query Data now rejects FOR XML/FOR JSON, table and query hints, and bare identity functions. Read the release and migration notes before upgrading, and CHANGELOG.md for the full dated history.

Everyday commands

sqlctx doctor check --mcp          # is everything ready?
sqlctx doctor update --host <host> # explicit update
sqlctx service status              # which supervisor owns the shared service

Working on this repository? Start at AGENTS.md; repair procedures are in docs/knowledge/.