Tai42 Dynamic Postgres

CI License: Apache 2.0

Schema-driven generator for safe, scoped PostgreSQL DML tools in FastMCP agent systems.

Point it at a PostgreSQL database and it introspects the schema and generates one typed MCP tool per DML operation per table — insert, select, update, delete, plus optional select_joined. The agent gets exactly those tools and nothing else: no raw SQL, no schema changes, no access to tables you didn't expose.

Why

Giving an agent database access usually means choosing between read-only access or risking that the agent runs arbitrary SQL and changes data or settings you never intended. This project gives you a controlled middle ground: a fixed set of generated, validated, parameterized tools scoped to the tables and operations you choose.

What it does

• Introspects your schema (tables, columns, types, foreign keys, pgvector).
• Generates a FastMCP tool per DML operation per table, each with a dedicated Pydantic input model derived from the table.
• Supports rich, injection-safe filtering (WhereFilter) and ordering, including vector KNN search.
• Excludes chosen columns from generated tool inputs (e.g. id, created_at).
• Exposes only the generated tools — never raw SQL or DDL.

Security model

[!IMPORTANT] The generated tool layer controls which operations exist; the PostgreSQL role you connect as controls what those operations can physically do. Connect as a dedicated least-privilege role, never a superuser. Read SECURITY.md before deploying.

Highlights, in full detail in SECURITY.md:

• Filter/order field names are validated against real columns and emitted only as quoted identifiers; values are always bound parameters. There is no path for SQL injection through field names.
• update/delete require a WHERE filter unless the server is started with --allow-unfiltered.
• The tool layer does not provide row-level scoping. Use PostgreSQL Row-Level Security and GRANT/REVOKE for that.

Installation

Requires Python 3.13+. Install from PyPI into the environment that runs the server, or run it without installing:

uv add tai42-dynamic-postgres-mcp
uvx --from tai42-dynamic-postgres-mcp tai42-postgres-mcp   # run it without installing

Or from source — clone this repo and add it as an editable dependency, or run the clone with uvx:

git clone https://github.com/tai42ai/tai-dynamic-postgres-mcp   # next to your app checkout
cd /path/to/your/app
uv add --editable ../tai-dynamic-postgres-mcp
uvx --from ../tai-dynamic-postgres-mcp tai42-postgres-mcp

Configuration

Connection and pooling are configured via environment variables:

Variable	Default	Description
PG_HOST	localhost	PostgreSQL host
PG_PORT	5432	PostgreSQL port
PG_DB	required	Database name (no default; startup fails if unset)
PG_USER	required	Database user, use a least-privilege role (no default)
PG_PASSWORD	required	Database password (no default; startup fails if unset)
PG_STATEMENT_TIMEOUT	30000	Per-connection statement_timeout in ms (0 disables)
PG_POOL_MIN_SIZE	1	Minimum pooled connections
PG_POOL_MAX_SIZE	10	Maximum pooled connections
PG_POOL_TIMEOUT	10	Pool acquire timeout (seconds)
PG_POOL_MAX_LIFETIME	300	Max connection lifetime (seconds)
TOOLS_DIR	~/.cache/tai42-dynamic-postgres-mcp/tools	Where generated tool files are written

CLI options

Flag	Default	Description
--overwrite / --no-overwrite	on	Regenerate the tool files on startup so they reflect the current schema. Pass --no-overwrite to reuse existing generated files
--readonly	off	Generate only select/select_joined tools
--allow-unfiltered	off	Allow update/delete to run without a WHERE filter (affects every row)
--select-joined a,b,c	—	Generate a joined select over the given tables (repeatable)
--ignore-insert-column	id	Column to exclude from insert inputs (repeatable)
--ignore-update-column	id	Column to exclude from update inputs (repeatable)
--ignore-select-column	—	Column to exclude from select output models (repeatable)
--ignore-select-joined-column	—	Column to exclude from joined select models (repeatable)
-t, --transport	stdio	stdio, http, sse, or streamable-http
--host	127.0.0.1	Bind host (HTTP/SSE transports only)
--port	8000	Bind port (HTTP/SSE transports only)

Usage with an MCP client

{
  "mcpServers": {
    "postgres": {
      "command": "uvx",
      "args": [
        "--from",
        "tai42-dynamic-postgres-mcp",
        "tai42-postgres-mcp",
        "--readonly"
      ],
      "env": {
        "PG_HOST": "localhost",
        "PG_PORT": "5432",
        "PG_DB": "dbname",
        "PG_USER": "agent",
        "PG_PASSWORD": "password"
      }
    }
  }
}

Generated tools

For a table public.orders you get (unless --readonly):

• select_public_orders(where, order_by, limit, offset)
• insert_public_orders(params, raise_on_conflict)
• update_public_orders(data, where)
• delete_public_orders(where)

Column types map to native Python: temporal columns to datetime/date/time, uuid to uuid.UUID, numeric/decimal to Decimal, json/jsonb to Any, and array columns to list[...]. insert returns the table's real primary key (a scalar list for a single-column key, a list of lists for a composite key, or the affected row count when the table has no primary key); columns with a database default are omittable. order_by items accept an optional nulls (FIRST/LAST); when unset PostgreSQL's default applies.

Filtering — WhereFilter

select, update, and delete accept a where argument. Field names must be real columns of the table; unknown fields are rejected.

// Simple field filters (implicitly ANDed)
{ "status": { "eq": "open" }, "total": { "gte": 100 } }

// Logical composition
{ "AND": [ { "status": { "eq": "open" } },
           { "OR": [ { "total": { "gt": 1000 } }, { "vip": { "eq": true } } ] } ] }

Supported operators: eq, ne, gt, gte, lt, lte, like, not_like, ilike, not_ilike, in, not_in, between, is_null, and knn (pgvector). Logical keys: AND, OR, NOT.

Vector search (pgvector)

When a column is a vector, filter or order by similarity:

{ "embedding": { "knn": { "query": [0.1, 0.2, 0.3],
                          "distance": "cosine",   // l2 | inner_product | cosine
                          "threshold": 0.5 } } }

Requires the pgvector extension enabled in the database.

Docker

docker build -t tai42-postgres-mcp .
docker run --rm -e PG_HOST=... -e PG_DB=... -e PG_USER=... -e PG_PASSWORD=... \
  tai42-postgres-mcp tai42-postgres-mcp --readonly

Development

See CONTRIBUTING.md.

uv venv --python 3.13
uv pip install --no-sources --group docs --editable ".[dev,test-integration]"
uv run --no-sync ruff check .
uv run --no-sync ruff format --check .
uv run --no-sync pyright
uv run --no-sync pytest --cov --cov-report=term-missing                 # unit tests
uv run --no-sync pytest -m integration  # CRUD tests against real Postgres (needs Docker)
uv run --no-sync mkdocs build --strict  # the docs site

License

Apache-2.0. See LICENSE and NOTICE.