- Reliability 8/20
- The README only shows a CI status badge and a coveralls badge as images/links, with no actual workflow file content or committed test evidence supplied, and no manifest was cached to confirm declared MCP tools match real behavior. The MCP mode is explicitly labeled 'EXPERIMENTAL' and described as offering 'a limited subset of tools for agents. Feedback is welcome,' signaling known incompleteness. Per the static calibration rule, without verifiable CI workflow content plus committed tests covering the server's key paths, the score cannot exceed 12, so 8 is assigned.
- Security and permissions 13/20
- Authentication uses standard OAuth, with credentials stored in `.clasprc.json` under the user's home directory; multi-user support (--user) and bring-your-own-OAuth-client guidance for better compliance are positive signals. However, the documented full scope list includes the very broad `https://www.googleapis.com/auth/cloud-platform`, which violates least privilege. Destructive operations like delete-script and delete-deployment prompt for confirmation by default (bypassable via --force, with the risk called out). The push command overwrites the entire remote script by default without an additional confirmation step, though this is prominently disclosed via a 'Warning' callout — still a real data-overwrite risk. The service account/ADC path is honestly labeled 'EXPERIMENTAL/NOT WORKING,' which is transparent but also confirms that pathway is unreliable. Overall, main risks are visible with partial confirmation and disclosure, but scope minimization and confirmation coverage are incomplete; this does not rise to red-line severity, hence 13.
- Maintenance 15/20
- The repository belongs to the google org, uses Apache-2.0, has 5778 stars and 53 open issues, and is not archived. A clear 2.x→3.x migration guide and recently added Gemini/Claude Code MCP extension support indicate active evolution. However, the supplied material shows no explicit security-response channel (e.g., SECURITY.md), release cadence, or ownership/CODEOWNERS evidence, so governance signals are incomplete, preventing a top score.
- Documentation 15/20
- The README is thorough: installation, authentication (including bring-your-own-credentials, allow-listing, service accounts), a full command reference with options and examples, `.clasp.json`/`.claspignore` configuration, troubleshooting (Node version, DEBUG logging, proxy usage), and a migration guide — well-layered with ample examples. However, documentation specific to the MCP mode being reviewed is very thin: a single descriptive sentence plus 'Options: N/A', with no listed tool names, parameters, return formats, or limits, and the manifest was not cached to verify actual tool behavior. This is a clear gap for the object under review, capping the score.
- Setup experience 12/20
- Clear npm global install steps are given, along with the need to manually enable the Apps Script API via an external GUI step, followed by `clasp login` for OAuth. For MCP specifically, three paths are documented: Claude Code plugin install, a `claude mcp add` CLI command, and referencing a `claude-mcp.json` config file — reasonable coverage of mainstream client configuration examples. However, the overall flow still depends on several manual/external steps (enabling the API, interactive OAuth login), and there is no execution evidence confirming reliability of this flow. Per the static calibration rule, the score cannot exceed 15, so 12 is assigned.