Loading ToolBench
Just let AI handle everything. Universal MCP server for Google Stitch with auto-setup, multi-CLI commands, professional design workflows, and AI image generation via Antigravity/Gemini 3 Pro.
stitch-mcp-auto has 5 tools with significant quality gaps. Tool names lack clear action verbs (generateImage, orchestrateDesign, removeBackground are somewhat acceptable; getAuthStatus and configureProject are weak). Descriptions are present but generic, ranging from 60-120 characters, below the 194-char baseline for production tools. Input schemas are visible and typed, but parameter descriptions are minimal (10-40 chars, far below the 72-char baseline). No output schema documentation visible in the code. Error handling is absent, no recovery guidance, no categorization of errors, no actionable error messages. The code shows basic OAuth/token management but no parameter validation, input sanitization, or security patterns like secret injection or permission gating. Composition is weak: orchestrateDesign conflates design creation, asset generation, and background removal into one tool, violating single-responsibility principle.
Set or verify the Google Cloud project ID
Generate images using Antigravity/Gemini 3 Pro with optional background removal
Check current authentication status including token expiry and Antigravity status
AI-powered design orchestration for UI/UX with automatic asset generation and background removal
Remove background from images with auto color detection or manual specification
orchestrateDesign violates single-responsibility principle: combines design brief processing, asset generation, and background removal. Should split into separate tools (create_design, generate_assets, remove_background) so agents can compose them independently.
Parameter descriptions are truncated and generic (10-40 chars each, baseline 72 chars). E.g., 'prompt' → 'Image generation prompt' lacks context on length limits, style guidance, or format requirements. 'backgroundColor' → 'Background color in hex format or auto-detection' should explicitly state the enum values and format pattern.
No output schema documentation. Users/agents cannot anticipate what fields generateImage returns (file path? URL? dimensions? format?). This forces trial-and-error and breaks downstream tool chaining.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | F | 39 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 26 | - | v1 |
No error handling or recovery guidance. Code contains try-catch blocks that catch exceptions but return raw error messages or silent failures (e.g., 'Could not fetch Antigravity project ID'). Agents have no actionable next steps (retry? ask user? fallback to alternative?)
No input validation or constraints declared. Parameters like 'prompt' (string) and 'backgroundColor' (hex or 'auto') have no length limits, format patterns, or enum declarations in the schema. This invites hallucinated values from LLMs.
Tool naming inconsistency: 'getAuthStatus' and 'configureProject' use camelCase verbs that are weak. 'getAuthStatus' should be 'check_auth_status' or 'get_authentication_status'. 'configureProject' is vague, does it create, update, or just validate? Use 'set_project_id' or 'update_project_config'.
OAuth credentials hardcoded in source code (GOOGLE_OAUTH_CLIENT_ID, ANTIGRAVITY_CLIENT_ID, etc.). These should be injected via environment variables. If this repo is public, credentials are compromised.
No permission gates or authorization checks. Any agent calling configureProject or generateImage can proceed without verifying the user's role or scope. Agents should be blocked on sensitive operations like project configuration.