MCP Database Doctor

MCP Database Doctor

Offline PostgreSQL diagnostics for AI coding agents. Review SQL, risky migrations, candidate indexes and supplied EXPLAIN JSON plans without database credentials or executing statements.

Status: 0.1.0 MVP. Static rules are heuristic: no_rules_triggered does not mean safe, performant or valid PostgreSQL. No telemetry, remote analysis, credentials or automatic fixes.

Install in your AI client

Works with any MCP client over stdio; no account or API key needed for the local server.

Claude Code

claude mcp add database-doctor -- npx -y mcp-database-doctor

Codex CLI

codex mcp add database-doctor -- npx -y mcp-database-doctor

Claude Desktop, Cursor, Windsurf, Cline, Gemini CLI — add to the client's MCP config (claude_desktop_config.json, ~/.cursor/mcp.json, ~/.codeium/windsurf/mcp_config.json, Cline MCP settings, ~/.gemini/settings.json):

{  "mcpServers": {    "database-doctor": {      "command": "npx",      "args": [        "-y",        "mcp-database-doctor"      ]    }  }}

VS Code / GitHub Copilot — .vscode/mcp.json:

{  "servers": {    "database-doctor": {      "type": "stdio",      "command": "npx",      "args": [        "-y",        "mcp-database-doctor"      ]    }  }}

Zed — settings.json:

{  "context_servers": {    "database-doctor": {      "command": "npx",      "args": [        "-y",        "mcp-database-doctor"      ]    }  }}

Tools

Reports include code, severity, message, recommendation, and statement numbers for SQL. No synthetic database health score.

Development

Node.js >=22.18 (Node 24 recommended).

npm installnpm run checknpm testnpm run buildnpm run test:integrationnpm pack --dry-run

Core tests run without installed dependencies using Node's native TypeScript stripping. Vitest integration tests exercise the official MCP client over stdio after build. CI runs both. Generate and commit a package-lock.json after the first successful dependency installation; current checkout does not include one because npm access was blocked in the implementation environment.

Claude Code / Cursor

Build locally first. Copy examples/mcp.json into your client's MCP configuration and replace the absolute checkout path:

{  "mcpServers": {    "database-doctor": {      "command": "node",      "args": ["/absolute/path/mcp-database-doctor/dist/server.js"]    }  }}

Claude Code CLI alternative:

claude mcp add database-doctor -- node /absolute/path/mcp-database-doctor/dist/server.js

Launch with npx -y mcp-database-doctor.

Suggested agent instruction: "Before proposing database changes, call check_migration. Review slow queries with analyze_query and supplied EXPLAIN JSON. Treat index DDL as candidates requiring workload validation."

Examples

check_migration({"sql":"BEGIN; CREATE INDEX CONCURRENTLY ON users(email); COMMIT;"}) identifies an invalid transaction context.

analyze_query({"sql":"SELECT * FROM users OFFSET 50000"}) flags projection and deep pagination.

explain_plan({"plan":"[{\"Plan\":{\"Node Type\":\"Seq Scan\",\"Plan Rows\":20000}}]"}) flags a scan for review; it does not claim an index is necessarily better.

For a real database, obtain plans yourself on a safe test environment: EXPLAIN (FORMAT JSON) SELECT .... EXPLAIN ANALYZE executes the statement; this server never runs it.

Limits and interpretation

• PostgreSQL-first, not a full SQL parser. Comments and string/dollar literals are masked. Quoted identifiers are masked to avoid keyword confusion.
• SQL inside stored procedures, DO blocks and dynamic strings is not analyzed. Complex CTEs, nested scopes, aliases, quoted/schema names and unusual DDL can produce false positives or missed findings.
• Index inference intentionally supports one unquoted table without joins/subqueries. No schema metadata, statistics, foreign-key index analysis or composite-index optimization. Existing indexes only suppress candidates when their supplied leading column matches; partial/expression indexes need manual review.
• Migration checks assume the supplied script defines transaction boundaries. If a migration framework wraps scripts externally, provide its BEGIN/COMMIT context when checking concurrent indexes.
• 100000 characters per SQL, 1 MB per plan string, 100 artifacts per report, 10000 plan nodes. These are analysis limits, not a substitute for host transport limits.
• Findings are review prompts. Absence of a finding is not authorization to run a migration.

Hosted path (quotas via control plane)

Local MCP stays free and offline. Quotas apply only on a hosted HTTP process that reserves units on mcp-control-plane before analysis. Build first (npm run build), then:

cp .env.example .env   # set CONTROL_PLANE_URLnpm run start:hosted   # default 127.0.0.1:3103

Requires Authorization: Bearer mcp_…. SQL and plans stay on the hosted host; control-plane sees only product, requestId, and units. No database credentials are accepted.

Commercial roadmap

Free: local query/migration/plan analysis. Pro later: history, before/after comparisons, CI policies and advanced recommendations. Team later: shared policies and centralized reports. Hosted quotas use the control-plane path above; local counters are not used for paid enforcement.

Validation status

• 53 core tests passed in the implementation environment.
• Typecheck/build/MCP stdio integration: configured, not run locally (npm registry returned HTTP 403).
• Real PostgreSQL and two actual AI hosts: pending; configuration files do not constitute host integration validation.

References

• https://ts.sdk.modelcontextprotocol.io/server
• https://www.postgresql.org/docs/current/sql-createindex.html
• https://www.postgresql.org/docs/current/using-explain.html

MIT license.