Loading ToolBench
An MCP server that proxies tool calls to the remote Stitch API (https://stitch.googleapis.com/mcp). It discovers tools dynamically from the remote API, registers them locally, and forwards calls with support for resilient screen generation with polling recovery.
Stitch MCP has 3 tools with mixed quality. Tool naming follows good verb-noun patterns (generate_, list_, check status), but parameter descriptions are sparse or missing entirely. Schemas are partially visible in code but lack comprehensive documentation. Two tools (generation_status, list_generations) have reasonable descriptions, but generate_screen_from_text lacks clarity on prerequisites and return structure. Input schema for generate_screen_from_text is inferred from handler code rather than explicitly registered in a structured way, parameters are validated in code but not formally exposed in schema. The server lacks output schema documentation, error recovery guidance, and parameter validation rules in descriptions.
Generates a screen from text (dynamically discovered from remote Stitch API; registered locally with resilient generation through tracker)
Check the status of a screen generation. Returns the generation's current state, elapsed time, and any errors.
List all in-flight and recent screen generations. Shows status of each generation including pending, polling, completed, and failed.
No output schema documented for any tool. LLMs cannot plan downstream tool calls or extract required fields without knowing return structure.
generate_screen_from_text description lacks prerequisites, return type, and state-modification clarity. Agents cannot determine if this is idempotent or has side effects.
Parameter descriptions missing or minimal. 'deviceType' and 'modelId' in generate_screen_from_text lack descriptions explaining their purpose, valid values, or optionality semantics.
Input schemas for generation_status and list_generations are minimal or inferred from code. No formal JSON Schema visible for list_generations (empty input) or structured generation_status schema documentation.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
Error handling does not guide recovery. Errors use generic messages (e.g., 'projectId is required') without suggesting next steps or alternative tools.
generate_screen_from_text is WRITE/destructive but description does not state this explicitly. Agents need to know which calls have irreversible side effects.
Tool composition risk: generate_screen_from_text is intercepted and routed through resilient generation tracker, but this routing logic is opaque to the MCP protocol layer. LLMs cannot infer the idempotence or retry semantics.
No pagination support or result limits documented for list_generations. If in-flight generations grow large, context window bloat is unaddressed.