- Reliability 11/20
- Only the README and repository description were readable; the server manifest is not cached, so the claim of 38 typed MCP tools cannot be checked against real behavior. The README asserts verification gates (typecheck, test, check:i18n, check:project-sync, check:skill, check:mcp, electron:check, build) and a release-desktop.yml workflow that boots the packaged server on three platforms and requires HTTP 200 from core routes — file-level evidence of intended reproducibility — yet no actual CI workflow content or committed test files covering the server's key paths are visible in this review. Under the static calibration rule, reliability must not exceed 12. Error and dependency handling is described concretely (90-second heartbeat timeout with retry of only pending/failed pages, resumable interrupted generation, cross-tab polling plus focus/visibility refresh) and dependencies are controllable (Node 20+, existing codex login, no extra API key). Missing visible test evidence and unverifiable tool parity drive the deduction. Scored 11.
- Security and permissions 12/20
- No red line is triggered: no malware, credential theft, covert exfiltration, or unconfirmed destructive defaults were found, and no install example contains real tokens or secrets. Credential handling is restrained — it reuses the existing codex login session and explicitly states no separate API key is stored and no .env is required; project data defaults to the repo's data/ directory with a warning not to commit personal project data or uploaded source material. However, permissions are only partially scoped: the MCP auto-start targets a loopback service (127.0.0.1:4311), but there is no allowlist or least-privilege statement covering outbound network use for web research and image generation; no explicit confirmation step is documented for side-effecting operations such as export, template/project deletion, or bulk rendering; and third-party community preview assets retain upstream CC BY 4.0 terms that the MIT license does not relicense, creating redistribution compliance risk. Incomplete scoping and confirmation warrant a deduction to 12.
- Maintenance 14/20
- The repository is not archived, has 881 stars and 7 open issues, ships an explicit MIT license, and includes SECURITY.md, CONTRIBUTING.md, CODE_OF_CONDUCT.md plus a private vulnerability reporting channel — a reasonably complete governance frame. The README shows a horizon-grouped roadmap (shipped / in progress / next / exploring) and a versioned release path (pnpm release, a manually triggered 'Release desktop apps' workflow producing a draft release and SHA256SUMS.txt), giving a clear version and update story. All of this is self-reported README evidence, though: static review cannot verify commit cadence, issue response latency, dependency update practice, or actual release history, so full marks are not justified. Scored 14.
- Documentation 17/20
- Documentation is the strongest dimension. Install (GitHub marketplace and local checkout), auth (codex login, no extra key), usage (first deck, file-sourced decks, modifying existing decks, mark-based edits, brand system), a grouped CLI/MCP tool table, an environment-variable table, on-disk data layout, local and desktop development commands, lifecycle commands, and troubleshooting (plugin invisible, app fails to start, stale tab, stuck rendering, non-editable PPTX text) plus a verification command list are all itemized. Limitations are disclosed fairly honestly: full-slide-image PPTX export with non-editable shapes is stated up front and listed on the roadmap, unnotarized desktop builds come with the exact bypass steps, and the media kit carries explicit safe-to-claim / do-not-claim guardrails. Deductions: per-tool parameter detail requires the referenced CAPABILITY_MATRIX.md, which is not visible here, and cost/latency is only characterized as roughly 4–5 minutes. Scored 17.
- Setup experience 13/20
- The install path is short and clear: two commands (codex plugin marketplace add, then codex plugin add), with a local clone plus pnpm dev alternative. Prerequisites are stated (Codex with plugin support, Node 20+, a prior codex login, Git), the port and URL are fixed at 127.0.0.1:4311, and environment variables have usable defaults. The dependency chain is still relatively heavy — Node 20 or newer, corepack/pnpm, and an existing authenticated Codex session — and the whole workflow is tightly bound to the Codex plugin system and the Codex in-app Browser; no configuration examples are given for other mainstream MCP clients (Claude Desktop, Cursor, etc.), and cross-client portability is unaddressed. With no verifiable in-file execution evidence, the static calibration caps setup at 15; whether a connection actually succeeds in a mainstream client cannot be confirmed, yielding 13.