What it does
SchemaBrain gives AI agents access to a database through a set of twelve MCP tools, none of which can write. The agent does not write SQL: queries are compiled from definitions the user controls. PII tags carry through the physical schema, joins and metrics, so a query that touches a blocked category is refused before the database is queried. Where the schema cannot answer a request, the server returns an unreachable-entity result with a suggested recovery tool rather than inventing a join. Every call, refusal and recovery is written to a SHA256-hashed append-only log, and an audit verify command exits non-zero if a past row has been altered.
It suits teams that want agents in Claude Desktop, Claude Code, Cursor, Windsurf or Cline to query a Postgres database under fixed rules, with a record of what was asked. Postgres is the only source supported at version 0.6.0 (beta); SQLite, Snowflake, BigQuery and MySQL connectors are listed as planned.
Setup uses a wizard, started with uvx schemabrain init, that asks for a Postgres connection string and an optional ANTHROPIC_API_KEY. A bundled demo runs with no API key and no Docker for the dashboard and showcase paths. On a user's own database, the wizard can use Docker and downloads an embedding model of about 67 MB on first run. Indexing a fresh 84-column schema with an LLM costs roughly $0.03, and re-indexing unchanged schemas costs nothing. Claude Desktop must be fully quit and relaunched to read the new configuration.
Limitations and requirements
Only Postgres is supported as a source today; the local store is SQLite. Audit logging is best-effort: if the disk is full or no writer is configured, a warning is logged and the query continues. Indexing a schema with an LLM on your own database requires an ANTHROPIC_API_KEY; without one, entity curation can run later. The first run downloads an embedding model of about 67 MB. Claude Desktop must be fully quit (not just its window closed) and restarted after configuration changes.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
SchemaBrain
The trust and intelligence layer between AI agents and your database.
v0.6.0As shown on registry.modelcontextprotocol.io when we last read it