- Reliability 8/20
- Evidence shows: the manifest is well-formed (valid schema reference, stdio transport, environment variables with isSecret flags, pinned npm version 0.3.4), and the declared tool surface (register, jmap_request, help) is consistent across the manifest and the README's description; the happy path is plausible. Deductions: the manifest declares package @atomicmail/mcp-modelcontextprotocol while the README instructs installing @atomicmail/mcp-github — a real contradiction; only the README and manifest were supplied, with no verifiable CI workflows, committed tests, or server source, so the init/tool-list handshake, error handling, and edge cases cannot be confirmed. Under the static calibration, missing execution evidence lowers the score, and the package contradiction pushes it below the 10 midpoint.
- Security and permissions 13/20
- Evidence shows: no red-line issue is evident — no real tokens in examples, no code-level exfiltration or covert data transfer, no irreversible destructive defaults; the intended data flow is disclosed (PoW auth service, JMAP API, local credential directory ~/.atomicmail/). Positives: ATOMIC_MAIL_API_KEY is flagged isSecret, credential files are written mode 0600, the README explicitly warns that inbound email is untrusted and scopes installation to the @atomicmail namespace. Deductions: the PoW protocol and credential lifecycle are only described at a high level; key rotation/revocation and a security contact channel are undocumented; no documented confirmation step exists for side-effecting operations such as sending mail (autonomous operation is the stated product goal, but an irreversible send lacks confirmation design); operational security processes are incomplete even though data-flow disclosure is clear.
- Maintenance 12/20
- Evidence shows: the repository carries an MIT license, is not archived, has 0 open issues, and the manifest and release badge indicate a versioned release path (0.3.4); the breadth of maintained integration surfaces (MCP, AgentSkill, LangChain, Dify, ClawHub) implies ongoing activity and clear version management. Deductions: commit frequency, dependency-update cadence, and issue-response behavior cannot be verified from the static material provided; the file-structure listing shows only CONTRIBUTING.md and LICENSE — no SECURITY.md or documented security-response channel; the CI/release badges are untrusted evidence and cannot directly prove activity.
- Documentation 13/20
- Evidence shows: the README is layered and substantive — quick install for multiple hosts, a docs-by-goal index table, example workflows, bundled JSON presets, a help tool with embedded cheatsheets, and disclosure of alpha status, the 100MB quota, and strict rate limits, plus an error-messages-with-hints design. Deductions: the install example contradicts the manifest's package name (@atomicmail/mcp-github vs @atomicmail/mcp-modelcontextprotocol), which can directly mislead users; pricing/cost beyond the free tier is not stated; the referenced docs/ files cannot be verified from the supplied material, so troubleshooting depth is only as strong as what the README itself demonstrates.
- Setup experience 10/20
- Evidence shows: as described, setup is short — paste an npx-based mcpServers JSON block into the host config and restart, after which the agent self-registers via PoW with no manual API key entry; all environment variables are optional and are documented with descriptions; stdio + npx has cross-platform potential. Deductions: the README install package name does not match the manifest's declared package, so users cannot tell which package is canonical — the path is ambiguous and fragile; setup depends on the external auth service's availability, the PoW protocol, and a Node.js 20+ environment; under the static calibration, the absence of verified CI workflows and committed tests caps this dimension below 15, and the package-name ambiguity further reduces the score.