android-mcp has 6 tools with basic HTTP transport via fastmcp. Tool naming follows verb_noun convention (list_devices, open_app, visit, current_window, press_button, wait), which is correct. However, descriptions are generic and parameter documentation is severely deficient. None of the 6 tools provide descriptions for their input parameters, only the tools themselves have descriptions. Input schemas exist but lack parameter-level documentation critical for LLM reasoning. Error handling is minimal (basic input validation on visit() and wait(), but no recovery guidance). No output schemas are documented. The codebase shows input validation (url stripping, seconds negativity check), but this is defensive rather than guidance-forward. Compared to production baselines (194 char avg description, 72 char avg param description, 100% A+ tools have param docs), this server underperforms significantly. The tools themselves are well-named but the supporting schema and documentation infrastructure is incomplete.

• Add parameter descriptions to all 6 tools. For open_app, document expected app name format (package name vs display name) and suggest calling a discovery tool if needed. For press_button, list valid button names explicitly or use an enum: ['power', 'volume_up', 'volume_down', 'home', 'back'].
• Document output schemas in tool docstrings. E.g., 'Returns {"devices": [{"serial": string, "status": string}], "connected": boolean}'. This tells LLMs exactly what fields exist and enables downstream planning.
• Expand visit() and wait() error messages with recovery guidance. Instead of raising ValueError, return {"status": "error", "reason": "URL must not be empty. Provide a non-empty HTTPS URL.", "retryable": true}. Enables LLM self-correction.
• Add a parameter description for 'seconds' in wait(): 'Number of seconds to wait (0 - 3600, integer). Use for synchronizing multi-step scenarios or rate-limiting adb commands.'
• Document what list_devices returns: total count, next_cursor (if paginated), and example device fields. State result cap if any.
• Consider adding a lookup/discovery tool (e.g., list_installed_apps, valid_hardware_buttons) to reduce hallucination in open_app and press_button.
• Add input validation error messages that teach the LLM the correct format. E.g., 'Invalid button name. Valid options: power, volume_up, volume_down, home, back, menu, search.'
• Ensure all response fields use consistent naming (avoid 'output' for raw adb output; name it 'adb_output' or 'command_output' to avoid confusion with tool output schema).