Loading ToolBench
Remote-network ADB MCP server. Drive Android devices over the network from any MCP-compatible agent (OpenClaw, Hermes, Claude Desktop, Claude Code, Cursor, Codex). Built for headless setups — droplets, CI, home labs, Tailscale meshes.
ADB MCP server has basic tool definitions but significant quality gaps. All 12 tools have names and descriptions, but descriptions are minimal (10-50 chars, well below the 194-char baseline). Input schemas are present but sparse, most tools have 1-2 parameters with type and description, but lack enums, ranges, constraints, and output schema documentation. No error handling guidance, no recovery hints, and no indication of which operations are idempotent or destructive (despite high-risk tools like uninstall and shell). Tool composition is reasonable (each does one thing), but parameter naming lacks consistency (e.g., 'x'/'y' vs 'x1'/'y1'/'x2'/'y2') and descriptions don't explain prerequisites or when to use each tool. Security concerns: shell tool accepts arbitrary commands with no validation or sandboxing hints. No mention of permission gates, audit trails, or secret injection patterns.
Get device information including model, Android version, and other properties
Install an APK on the device
Send a key event to the device
Pull a file from the device to the host
Push a file from the host to the device
Capture the current screen of the Android device and return as base64-encoded PNG
Execute a shell command on the Android device
Swipe from one point to another on the device screen
Descriptions are critically short (10-50 chars vs. 194-char baseline). Most lack WHAT/WHEN/WHY context. E.g., 'shell' says 'Execute a shell command' but doesn't explain prerequisites, risks, or when to use it instead of higher-level tools.
No output schemas documented. LLMs cannot plan downstream calls or extract structured data. E.g., screencap returns base64 PNG but no schema states this; ui_dump returns XML but no schema describes the structure.
shell tool accepts arbitrary commands with no validation, sandboxing, or permission hints. High security risk, LLMs could be tricked into executing destructive commands. No error guidance for command failures.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
Tap at the specified coordinates on the device screen
Type text on the device
Dump the current UI hierarchy as XML from UIAutomator
Uninstall an app from the device
Destructive tools (uninstall, shell, push, pull) lack confirmation/dry-run patterns and error recovery guidance. No indication of which operations are idempotent or irreversible.
Parameter descriptions lack constraints. E.g., 'x' and 'y' for tap/swipe don't specify pixel ranges, coordinate system, or screen bounds. 'duration' for swipe has no min/max. LLMs will guess invalid values.