Loading ToolBench
A keyboard executor skill for mobile agentic systems that processes keyboard input to execute contextual actions such as calendar invite modals, one-click payments, and smart reply synthesis.
This server has severe structural issues that prevent it from being production-ready. The code provided is a mock Python client library, NOT an MCP server. There is no MCP protocol implementation visible: no transport layer (HTTP, SSE, or STDIO), no tool registration via MCP-compliant schemas, no integration with the Model Context Protocol framework. The single tool 'execute_from_keyboard' has a minimal description, no parameter descriptions, undocumented output structure, and hardcoded intent confidence values. The implementation is a thin wrapper around keyword matching with no error handling, validation, or structured output.
Processes keyboard input in a given app context and returns an executed action with intent confidence. Recognizes keywords related to scheduling, meetings, payments, transfers, and defaults to smart reply synthesis.
NOT AN MCP SERVER, Code is a mock Python client library with no MCP protocol implementation. No transport layer (HTTP, SSE, or STDIO) is visible. No MCP Server implementation, schema registration, or protocol compliance.
Tool description is generic and does not explain WHEN to use this tool. Description: 'Processes keyboard input in a given app context and returns an executed action with intent confidence.' Missing: what confidence threshold is used, what app contexts are valid, what smart reply synthesis means, when to prefer this tool over alternatives.
Parameter 'keyboard_input' has a type but no description. Parameter 'app_context' has a type and default but no description explaining valid values or format.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 23 | 2026-07-28+ | v2 |
Output schema is undocumented. The tool returns a dict with keys 'executed_action' (string) and 'intent_confidence' (float 0.94 hardcoded), but there is no formal schema definition or documentation of what these fields mean, what their ranges are, or what structure downstream tools should expect.
Intent confidence is hardcoded to 0.94 regardless of input. This is not a real confidence metric, it is a lie to the LLM. The LLM cannot trust a constant confidence value and cannot distinguish high-confidence predictions from low-confidence ones. Confidence must reflect actual model uncertainty.
No validation of input. If 'app_context' is 'invalid_context', the tool silently accepts it and returns a response. There is no error handling, no list of valid contexts, no guidance to the LLM on what went wrong.
Keyword matching is naive and brittle. The tool checks if 'schedule' is in lowercase text to trigger calendar actions. This will fail on variations ('scheduling', 'scheduled'), synonyms ('book', 'reserve'), and will hallucinate actions when these words appear in unrelated contexts (e.g., 'Do not schedule this payment').
No documentation of what app_context values are valid (messaging, email, payments, calendar?). Parameter description is missing. LLM will guess or hallucinate values.
No error handling at all. If input is malformed or execution fails, there is no error response, recovery guidance, or actionable message for the LLM. A tool that silently fails or returns nonsense values breaks agent planning.