- Reliability 7/20
- Supplied material is README plus repo metadata only: no manifest, no source, no CI workflows, no committed tests. The README describes a built-in MCP server at /mcp/ over Streamable HTTP with 15 tools whose list is coherent with the documented REST surface, and it even documents error-shaped behavior (deprecation headers, 400/422 rejection of conflicting targeting), so the happy path reads plausibly. However, nothing verifies MCP initialization, the tool-list handshake, or actual tool behavior; the README itself shows client-side friction (Claude Desktop needs an mcp-remote bridge; certificate trust varies per client). Static calibration caps reliability at 12 without execution evidence; with tests/CI absent and per the conservative round-down rule, scored 7.
- Security and permissions 14/20
- No red line triggered: API-key bearer authentication, server bound to 127.0.0.1, examples use only placeholders, the plain-HTTP endpoint is opt-in with an explicit insecurity warning, and the delete tool defaults to trash, mitigating destructive side effects. Deductions: no per-operation confirmation flow is documented for dangerous actions, no security policy or disclosure channel is visible, the extension API widens the plugin trust boundary by letting other plugins register routes, and the self-signed certificate plus HTTP fallback, while disclosed, is inherently weak. Least-privilege binding and data boundaries are mostly clear, so above 10; still short of 'explicit confirmation + full data-flow disclosure', hence 14.
- Maintenance 13/20
- Evidence: MIT license, repo not archived, very low open-issue count (2), README shows active versioning (deprecation schedule targeting 6.0), a published typed extension package (obsidian-local-rest-api), a contributing guide, and a known-extension ecosystem. Deductions: commit history, release cadence, dependency-update process, and a security-response channel are not verifiable from the supplied materials. Solid but not full marks: 13.
- Documentation 16/20
- The README is layered and strong: quick starts for the REST API and four MCP client types, a full endpoint table, MCP tool/resource tables, auth and certificate instructions, deprecation and migration notes, plus links to interactive API docs and a wiki. Deductions: no parameter-level schema for MCP tools in the README (only names and one-line descriptions), troubleshooting content is sparse (mostly certificate warnings), and no explicit limits/resource disclosure. Not a perfect 20, so 16.
- Setup experience 14/20
- The mainstream path is few-step and concrete: install and enable the Obsidian plugin, copy the API key from settings, paste ready-made client configuration (full JSON for Claude Code/Cursor, an mcp-remote command for Claude Desktop); certificate download/trust and the HTTP fallback are explained, with macOS/Windows paths given. Deductions: Claude Desktop requires Node.js and an mcp-remote bridge, self-signed certificate trust varies by client and adds friction, and static calibration caps setup at 15 without execution evidence (no CI/tests confirm these configs work), so 14.