PostgreSQL查询服务

Electric Elephant

Electric Elephant is a token-efficient MCP server for exploring and querying PostgreSQL from MCP-capable clients. It is not a generic SQL bridge: only PostgreSQL is supported (not MySQL, SQLite, SQL Server, Oracle, or other engines).

Schema scoping: Every row-returning tool (execute_sql, search_objects, explain_plan, custom tools) requires a per-call schema argument, and SQL is confined to that single schema — cross-schema and system-catalog (information_schema, pg_catalog) references are rejected before any other check, and the session search_path is pinned with SET LOCAL.

Health and PII data: Health/clinical data (HL7v2, FHIR, LOINC, SNOMED, medical fields), personal & sensitive data (names, email, government/financial IDs, DOB/age, address/location, biometric/device IDs, credentials/secrets), and wildcard projections (SELECT *, table.*) are hard-excluded on execute_sql and custom tools — they can never be returned, and allow_access_to_pii_data does not unblock them. The only field the opt-in unblocks is the user's mobile/phone number (the Helium username). These are heuristic, name-based checks (best-effort, can false-positive/negative) — a safeguard, not certification or a substitute for database permissions, row-level security, legal review, or your own data policies. See docs/tools/execute-sql.mdx.

Repository: github.com/ajgreyling/electric-elephant

Upstream Sync Status

Electric Elephant is a fork of dbhub. PostgreSQL-relevant upstream fixes are synced through dbhub commit 72adfdce530ebaf2d7e6df12de5ecde0d174cf4f (2026-04-21), on top of upstream release line v0.21.2.

Backported upstream commits:

• f319114033279532aff2ce9aaef2ce84b127a21f (PostgreSQL getTableComment() view support)
• ce2621d83d78d9ab8b363664c955584cb59ee049 (graceful skip on transitive MODULE_NOT_FOUND)
• 30d8007998503defc05d5198bcbd9130c609ee41 (HTTP DNS rebinding protection)
• f13fad459d1ac9f7837fc39e37941247bd6d0c6d (PostgreSQL SSL verify-ca/verify-full + sslrootcert)
• f35144b87f4394dd7a36416b7a459bfd710b61f4 (SSL documentation updates for verify-ca / verify-full)
• 72adfdce530ebaf2d7e6df12de5ecde0d174cf4f (source description surfaced in MCP tool descriptions)

Purpose

• Expose PostgreSQL through MCP tools (execute_sql, search_objects, query_insights, schema_diff, observability helpers, and related wiring).
• PostgreSQL-only: no connectors or compatibility layers for other SQL databases.
• Provide safe defaults (read-only unless explicitly enabled for destructive SQL).
• Single-schema scope: every row-returning tool requires a per-call schema and confines SQL to it; cross-schema and system-catalog references are rejected (SCHEMA_SCOPE_VIOLATION).
• Hard-exclude health & PII (best effort): on execute_sql and custom tools, health/clinical data (HL7v2/FHIR/LOINC/SNOMED-style + medical heuristics), personal & sensitive data (names, email, government/financial IDs, DOB/age, address/location, biometric/device IDs, credentials/secrets), and wildcard projections are always blocked — name-based heuristics can false positive/negative. allow_access_to_pii_data does not unblock these; the only field it unblocks is the user's mobile/phone number (the Helium username). Opt-in: TOML allow_access_to_pii_data, env ALLOW_ACCESS_TO_PII_DATA, or single-DSN CLI bare --allow-access-to-pii-data (or =true / 1 / yes). Destructive SQL in single-DSN mode: same pattern with --allow-destructive-sql. See docs/tools/execute-sql.mdx, docs/config/command-line.mdx, and CLAUDE.md.

Repository Landmarks

• src/index.ts - entrypoint and startup path.
• src/server.ts - HTTP MCP transport wiring.
• src/connectors/ - database connector implementations.
• src/tools/ - MCP tool handlers (execute_sql, search_objects, query_insights, schema_diff, etc.).
• src/config/ - TOML/config loading and validation.
• frontend/ - local web workbench UI.
• CLAUDE.md - architecture and development conventions.

Installation

For end users running Electric Elephant as an MCP server:

NPM (npx):

npx electric-elephant --transport http --port 8080 --dsn "postgres://postgres:postgres@localhost:5432/postgres"

Docker:

docker run --rm --init \
  --name electric-elephant \
  --publish 8080:8080 \
  electric-elephant \
  --transport http \
  --port 8080 \
  --dsn "postgres://postgres:postgres@host.docker.internal:5432/postgres"

See docs/installation.mdx and docs/quickstart.mdx for full client setup instructions.

Development Quick Start

pnpm install
pnpm run dev

Build and test:

pnpm run build
pnpm test

Workbench

Electric Elephant includes a built-in web Workbench for running tools and inspecting request traces.

• Start server with HTTP transport (examples above), then open http://localhost:8080
• Workbench UI: /
• MCP endpoint: /mcp

More details: docs/workbench/overview.mdx

MCP Request Flow

flowchart LR
    A[MCP Client] --> B[Transport: stdio or HTTP]
    B --> C[Tool Router]
    C --> D{Tool}
    D -->|execute_sql search_objects query_insights schema_diff ...| E[Connector Manager]
    E --> F[PostgreSQL connector]
    F --> G[(PostgreSQL)]
    G --> F --> E --> C --> A

All built-ins (including read-only diagnostics such as explain_plan, diagnose_locks, and replication_status) route through the connector manager to the same PostgreSQL connection pool for the selected source.

Built-in MCP tools

These tools are enabled by default per [[sources]] entry unless you whitelist a subset with [[tools]] in dbhub.toml. With multiple sources, names are suffixed with the source id (for example execute_sql_prod_pg).

User-defined [[tools]] entries add custom parameterized SQL tools. See docs/tools/overview.mdx and dbhub.toml.example.

Query Execution State Machine

stateDiagram-v2
    [*] --> RequestReceived
    RequestReceived --> ValidatingInput
    ValidatingInput --> SelectingSource
    SelectingSource --> Executing
    Executing --> FormattingResponse
    FormattingResponse --> Completed
    Executing --> Failed
    ValidatingInput --> Failed
    Failed --> [*]
    Completed --> [*]

Human + AI Agent Onboarding Checklist

1. Read CLAUDE.md before editing connectors/tools.
2. Prefer tool-level changes in src/tools/ over transport-layer changes.
3. Keep source_id routing behavior backward compatible.
4. When changing execute_sql, preserve schema-scope enforcement (sql-schema-scope.ts, SCHEMA_SCOPE_VIOLATION) and PII/health guard semantics (pii-sql-guard.ts, pii-heuristics.ts, PII_ACCESS_VIOLATION) — never let allow_access_to_pii_data unblock anything but the mobile number.
5. Run relevant tests (pnpm test, or targeted connector/integration tests).

Tool Schema Examples

execute_sql input (schema is required; list explicit non-sensitive columns — SELECT *, health/clinical data, and personal identifiers are always rejected and cannot be enabled; only the mobile number is unblockable via allow_access_to_pii_data):

{
  "schema": "public",
  "sql": "SELECT id, status FROM users LIMIT 10;"
}

search_objects input:

{
  "object_type": "column",
  "schema": "public",
  "table": "users",
  "pattern": "%_id",
  "detail_level": "summary",
  "limit": 50
}

Related Docs

• docs/tools/overview.mdx — all MCP tools and TOML whitelisting.
• docs/tools/execute-sql.mdx — execute_sql, read-only mode, PII guard.
• docs/config/command-line.mdx — CLI flags including --allow-access-to-pii-data (single-DSN).
• docs/tools/search-objects.mdx — search_objects patterns and detail levels.
• docs/tools/query-insights.mdx — query_insights and pg_stat_statements.
• docs/tools/schema-diff.mdx — schema_diff between two sources.
• docs/tools/custom-tools.mdx — parameterized custom tools.
• dbhub.toml.example — multi-source and tool configuration examples.
• Mintlify site config: docs/docs.json.