- Reliability 8/20
- The README describes a unified stateless Streamable HTTP handler (/mcp with an /sse compatibility alias) and explains that OAuth, credentials, account selection, and product state live in application-layer Durable Objects — the architecture description is coherent. However, server.json declares no tools at all, so declared tool behavior cannot be checked against real behavior, and no CI workflow files or committed tests are present in the provided material as verifiable execution evidence. Per the static calibration rule, reliability is capped at 12 without CI+test evidence, and the 16 remote endpoints with a mixed optional-Bearer/OAuth auth story add further uncertainty. Scored 8: the happy path is plausible from the description, but error handling, edge cases, and reproducibility lack supporting evidence.
- Security and permissions 11/20
- Most remote endpoints authenticate via an optional Authorization Bearer header or OAuth, with the key field marked is_secret — a reasonable credential-handling pattern with no hardcoded tokens or obvious leakage. The README shows an example of creating a scoped, minimal-permission API token for a specific service (Browser Rendering), indicating some least-privilege awareness. However, the material does not disclose whether destructive or irreversible operations (DNS changes, Workers deployments, container spin-up, audit log export, etc.) require confirmation, and the Bearer token grants are described only at a broad User/Account level without per-domain scoping detail across the 16 product areas. No red-line issue is present (no malware, credential theft, or fabricated results), so it does not fall into the 0-4 band, but incomplete evidence on confirmation flows and permission scoping caps it at 11.
- Maintenance 14/20
- The repository uses the Apache-2.0 license, the README references a CONTRIBUTING.md indicating a contribution process, and with 4,028 stars, 53 open issues, no archival, and an 'active' status field in the manifest, it shows clear signs of active maintenance. However, the provided material contains no concrete release history, commit cadence, dependency-update policy, or dedicated security-response channel, leaving governance details incomplete. Scored 14: active but without enough evidence of versioning/security-response process for a higher score.
- Documentation 13/20
- The README provides a service table with URLs and subfolder links, connection instructions, an OpenAI Responses API integration example (with a permission-scope screenshot), a troubleshooting section (context-length interruptions), and a note on paid features — a reasonably layered set of docs. However, it lacks per-tool parameter/response documentation, rate-limit or cost specifics, and repeats the same generic Authorization description across all services rather than detailing per-domain usage examples or known limitations. Scored 13.
- Setup experience 13/20
- Because all servers are hosted remote Streamable HTTP endpoints, users simply plug a URL into any MCP client with remote-server support (or the Cloudflare AI Playground) — no local runtime is required, which is a genuine ease-of-use advantage. Per the static calibration rule, setup is capped at 15 absent verifiable CI/test execution evidence; additionally, the material does not detail the concrete OAuth authorization interaction steps or local development setup (it only references an external CONTRIBUTING.md without expanding on it). Scored 13.