DBHub Database MCP Gateway

The evidence shows a two-tool default surface plus read-only mode, row limits, query timeouts, transactions, SSH tunneling, and TLS, providing substantive least-privilege, external-effect, and transport protections. No user-confirmation mechanism for dangerous SQL is described, however, and it is unclear whether read-only mode is enabled by default. The client–DBHub–database diagram and tool descriptions make the primary data path visible, but telemetry, logging, Workbench trace retention, and credential handling are not fully disclosed. The installation example embeds a password in a command-line DSN, risking exposure through shell history or process listings. Dependencies use version ranges, while no lockfile, dependency audit, vulnerability scan, or update policy appears in the supplied evidence. Transactions offer limited rollback, but recovery for committed writes or non-transactional DDL is unexplained. Bytebase, the repository, and MIT copyright are attributed, yet package.json has an empty author field and publisher identity is unverified, so attribution is not comprehensive.

The README, package metadata, and publishing workflow generally agree on supported databases, Node requirements, naming, and distribution. The README's “zero-dependency” claim directly conflicts with the runtime and optional dependencies in package.json, reducing self-consistency. Node and pnpm versions, database drivers, and npm/Docker delivery routes provide a reasonable availability basis, but driver selection and missing-driver behavior are not explained. A health-check tool, Debug link, and test script names are present, but the supplied material contains no concrete error-message examples, diagnostic taxonomy, or recovery guidance, leaving failure messaging thin.

Audiences and scenarios are thoroughly described across local development, non-technical access, multi-database consolidation, and production read-only troubleshooting. Supported databases, default and opt-in tools, custom tools, transports, and multi-connection configuration establish useful capability boundaries, but write boundaries, database-specific differences, and security defaults remain incomplete. A small default tool set and explicit opt-ins support precise selection, though tool schemas, invocation conditions, and ambiguity handling are absent. Node >=22.5, stdio/HTTP, Docker, MCP Bundle, plugin delivery, and multi-architecture images support varied environments, but OS coverage, database-version compatibility, and deployment constraints are not documented here.

The README has a coherent progression through use cases, databases, tools, Workbench, installation, and development, with links to focused documentation; important details nevertheless depend on external pages not supplied. Installation and development commands, the Node requirement, and several packaging options are clear, but the use of latest and an inline credential is risky, and upgrade, uninstall, and production-hardening notes are absent. DBHub naming, package and MCP identifiers, and tool names are stable across the evidence, justifying full naming credit. Examples are limited to startup commands and use-case summaries, with no result walkthrough or FAQ. Node constraints, opt-in tools, and read-only use are mentioned, but there is no consolidated limitations section. MIT metadata exactly matches a complete LICENSE, earning full credit. Version 1.2.0, a synchronization script, and Docker version-tag automation show an update path, but no changelog, release notes, or compatibility policy is supplied. Bytebase attribution, a contributor route, and repository ownership indicate maintenance responsibility, while the empty author field, lack of named maintainers/support commitments, and unknown publisher identity prevent full credit.

The two default tools, progressive object search, parameterized custom SQL, Workbench, and unified multi-database access indicate outputs intended for immediate use by MCP clients and humans. Actual response schemas, errors, pagination, and result examples are absent, limiting confidence in output usability. Consolidating several database backends behind a small tool surface offers clear marginal value, although feature tradeoffs versus alternatives are not developed. The README quantifies a 1.4k-token footprint and a 13–14x comparison, but provides no measurement method, version baseline, or reproducible artifact, and does not quantify operational or security costs; cost-benefit handling is therefore adequate rather than thorough.

Many basic claims trace to the README, package.json, LICENSE, and publishing workflow, including database coverage, tool count, Node requirements, versioning, and image publication. Token counts, competitor comparisons, the “official example” statement, and claimed safety effects lack supporting measurements or source artifacts in the provided files. Multiple files corroborate naming, version, licensing, runtime requirements, and delivery paths, but major performance and security claims are largely README-only. Factual descriptions, comparative conclusions, and promotional language are intermingled without clearly separating implementation evidence, recommended usage, and measured results, so fact–inference separation is weak.