- Reliability 9/20
- From the static artifacts only (README, server.json), the manifest is internally consistent: package identifier appium-mcp, version 1.94.0, npm registry type and stdio transport are coherent, and the README documents session lifecycle, disconnect cleanup policies (delete_all/skip) and normalized error codes, which suggests diagnosable failures. However, no CI workflows, committed tests, or reproducible logs are visible, and the tool table is truncated, so declared tools cannot be verified against actual registrations. Static calibration caps this dimension, and the score is pulled down by missing executable evidence and unproven edge-case/error handling.
- Security and permissions 12/20
- No red line is triggered: no malware, credential theft, covert exfiltration, or default destructive remote commands. The README shows real security awareness: AI vision and documentation tools are opt-in and env-gated, remoteServerUrl supports REMOTE_SERVER_URL_ALLOW_REGEX allowlisting, query strings/fragments are rejected, redirects are not followed for capability discovery, URL userinfo is redacted from logs, and persisted session files use 0600 permissions. Still, defaults are permissive: any HTTP(S) destination is accepted when the regex is unset, and client disconnect deletes all MCP-owned Appium sessions without confirmation, a potentially destructive side effect. remoteServerUrl and capabilities are injectable via tool arguments with no mandatory confirmation step, and there is no process-level least-privilege or sandbox constraint, so the score lands mid-range.
- Maintenance 11/20
- The license is a clear Apache-2.0, and versioning at 1.94.0 with frequent-looking iteration suggests active release cadence; 477 stars, 23 open issues and a non-archived repo are activity signals (treated only as discovery, not scoring credit). The dependency stack (appium-uiautomator2-driver, appium-xcuitest-driver, FastMCP) and README treatment of OpenTelemetry, plugin API and an optional docs package indicate some governance thinking. But no commit history, release notes, dependency-update policy, security response channel, or explicit maintainer ownership is visible in the static material, so verifiable maintenance evidence is incomplete and the score is capped below full marks.
- Documentation 14/20
- Coverage is notably broad: installation (one-click Cursor, Gemini CLI, Claude Code), a full environment variable table, capabilities samples, the general platform mode, multi-session and sessionId guidance, disconnect-cleanup tradeoffs, screen recording, AI vision model benchmarks and coordinate handling, NO_UI/MCP Apps performance tradeoffs, a plugin API with lifecycle, and a remote server trust model with regex allowlist examples. No real tokens or secrets appear in the examples. Deductions: the tool table is truncated in the supplied material, several parameters and response shapes lack complete examples, troubleshooting is thin, and benchmark claims such as 100% accuracy cannot be independently verified through static review.
- Setup experience 11/20
- The installation path is clear: Node 22+, JDK, Android SDK and Xcode prerequisites are stated, the npx appium-mcp@latest stdio config is reusable across mainstream clients, and Cursor deep links plus Gemini CLI and Claude Code CLI commands are provided. In practice, however, it depends on heavy manual environment preparation (local Android SDK/ANDROID_HOME, Xcode and device signing, real-device provisioning profiles, ffmpeg), which is fragile, and general platform mode requires remoteServerUrl. End-to-end reproducible integration examples and common failure troubleshooting are absent, so the score stays below the static calibration ceiling.