- Reliability 8/20
- Only a README was supplied—no source code, test files, or CI workflows were available to verify that the 31 declared tools actually behave as described or that the MCP init/tool-list handshake completes cleanly. The README's per-tool input/output descriptions are detailed and internally consistent, suggesting a reasonably mature implementation, but there is no reproducible execution evidence (no committed tests, no CI logs). Per the static calibration rule, reliability is capped at 12 without such evidence, so a modest score reflects 'plausible but unverified.'
- Security and permissions 9/20
- Credentials can be passed via CLI flags (--clientSecret/--accessToken) or a .env file; CLI flags carry the usual risk of exposure via shell history or process listings, though all documented examples use placeholders rather than real secrets. The tool surface includes several irreversible operations (delete-product, delete-customer, order-cancel explicitly marked 'Irreversible', refund-create, order-mark-as-paid, complete-draft-order), but these require explicit required parameters (e.g., reason, restock) rather than executing with silent defaults, so this falls short of the full red-line 'irreversible action defaults without confirmation' scenario. Still, no built-in server-side confirmation step is documented, and least-privilege scoping is left entirely to the user's OAuth app configuration, which limits the score.
- Maintenance 10/20
- The repository is MIT-licensed, not archived, and shows 235 stars with 10 open issues, indicating some ongoing activity, but no commit history, release/version cadence, changelog, or security-response channel was provided in the materials. Absent evidence is treated as a deduction rather than assumed positive, yielding a mid-range score.
- Documentation 16/20
- The README is thorough, covering both authentication methods (OAuth client credentials and legacy static tokens), Claude Desktop/Claude Code configuration examples, an environment-variable alternative, detailed per-tool parameter documentation for all 31 tools, pagination/sorting/filtering conventions, an order query filter reference table, and a troubleshooting note about package-name confusion. It falls short of full marks because it omits API cost/rate-limit disclosure, does not map required OAuth scopes to individual tools, and lacks systematic error-code/failure-mode documentation.
- Setup experience 13/20
- Setup steps are clear: npx-based execution, optional global install, JSON config examples for Claude Desktop, a Claude Code CLI command, and an equivalent .env-based flow, with both new and legacy auth options documented — low friction for mainstream client users. However, per the static calibration rule, setup cannot exceed 15 without verifiable execution evidence (CI plus committed tests covering key paths), so the score is high within that cap but not higher.