mcp-server-pgvector

An MCP server that gives LLM agents first-class access to pgvector-backed embedding tables in PostgreSQL: similarity search, hybrid (vector + full-text) search, upserts, and HNSW/IVFFlat index management.

Generic Postgres MCP servers expose raw SQL or schema introspection; this one speaks pgvector specifically — nearest-neighbor search, distance metrics, and ANN index tuning are first-class tools, not something the model has to hand-write SQL for.

Tools

Safety

• Every table/column name is validated against information_schema / pg_catalog before being interpolated into SQL — an LLM can only ever reference identifiers that already exist. Values are always bound parameters.
• Metadata filters are a closed {column, op, value} allowlist, not a raw SQL fragment.
• Set MCP_PGVECTOR_READ_ONLY=true to disable upsert_embedding and create_vector_index, leaving only read/search tools available — useful when pointing the server at a production database.
• Every query runs with a per-command timeout (MCP_PGVECTOR_COMMAND_TIMEOUT_SECONDS, default 30s) so one expensive query can't occupy a pool connection — and stall every other caller — indefinitely. Set it to 0 to disable.

Installation

uvx mcp-server-pgvector

Or with pip:

pip install mcp-server-pgvector
python -m mcp_server_pgvector

Configuration

The server reads its connection string from DATABASE_URL (or PGVECTOR_DATABASE_URL):

{
  "mcpServers": {
    "pgvector": {
      "command": "uvx",
      "args": ["mcp-server-pgvector"],
      "env": {
        "DATABASE_URL": "postgresql://user:password@localhost:5432/mydb",
        "MCP_PGVECTOR_READ_ONLY": "false",
        "MCP_PGVECTOR_COMMAND_TIMEOUT_SECONDS": "30"
      }
    }
  }
}

Production readiness

Covered:

• Identifier-safe SQL (every table/column checked against pg_catalog before use) and a closed filter-operator allowlist — no path from tool arguments to raw SQL.
• Per-query timeout, so one runaway query can't monopolize the (small, 5-connection) pool.
• 60+ tests, including dimension-mismatch and injection-attempt regressions, run in CI on every push/PR against a real pgvector container across Python 3.10–3.13. A separate CI job builds the package and runs twine check on the result.
• Connection failures surface as plain ConnectionRefusedError/asyncpg exceptions — verified these don't leak the DSN's credentials into error text.

Known limitations, honestly:

• No per-tool authorization — access control is whatever the Postgres role in DATABASE_URL can do. If you need different agents to have different permissions, give them different connection strings backed by different Postgres roles, not different server instances of this same DSN.
• hybrid_search's full-text side is hardcoded to Postgres's 'english' text search configuration; there's no parameter to change it yet.
• No structured logging — failures are exceptions surfaced through the MCP error channel, not written to a log you can tail. Fine for a single-user desktop MCP client, a real gap if you're running this as a shared service.
• The connection pool is fixed at 1–5 connections and isn't configurable via environment variable yet.

Development

uv sync --dev

# Bring up an isolated pgvector instance for local testing
docker compose -f docker-compose.dev.yml up -d

export DATABASE_URL=postgresql://postgres:postgres@localhost:5434/postgres
uv run pytest

uv run ruff check .
uv run pyright

Contributing

See CONTRIBUTING.md. See CHANGELOG.md for release history.

License

MIT — see LICENSE.