Static source inference · medium confidence · detected: Logging
Deprecated protocol patterns detected
Summary
stitch-pro-mcp demonstrates solid schema coverage with Zod-based input validation and mostly clear tool naming. However, output schemas are undocumented (no return types visible in source), descriptions are sparse for many tools, and error handling is absent from the provided code. The server has 11 well-named tools with action verbs (list_, get_, generate_, create_, apply_, convert_, extract_) and typed parameters via Zod schemas. Parameter descriptions are generally present and adequate (e.g., 'Description of the page to generate', 'Target device'). However, critical gaps include: (1) no visible output schema documentation for any tool (Pattern:response-shaper violation), (2) tool descriptions averaging ~60 chars, below the 194-char baseline for A+ tools, (3) missing error handling implementation or recovery guidance, (4) no evidence of permission checks or audit trails despite write operations. The pipeline architecture with processors (AccessibilityFixProcessor, ResponsiveInjectProcessor, etc.) suggests sophisticated internal logic, but this complexity is not reflected in tool documentation or error messages visible to agents.
Tools (11)
sp_a11ywritesource verified78/100
Run WCAG 2.1 AA accessibility audit on HTML and auto-fix issues (contrast, semantics, ARIA, touch targets)
sp_apply_design_systemwrite50/100
Apply a design system to existing HTML
sp_convert_htmlreversible50/100
Convert HTML/Tailwind to a framework (React, Vue, Svelte) with optional component library mapping
sp_create_design_systemwrite50/100
Create a new design system with brand colors, typography, and spacing tokens
sp_extract_componentsread only50/100
Extract reusable components from HTML and map to a component library (shadcn, Radix UI, Material-UI)
sp_generate_flowwriteauth50/100
Generate a multi-screen user flow in Stitch
sp_generate_pagewriteauth50/100
Generate a single page UI design in Stitch with optional design system, accessibility, and responsive features
No output schemas documented for any tool. LLMs cannot plan downstream tool calls or extract required data without knowing what fields to expect in responses.
Tool descriptions are incomplete and lack guidance on when to use each tool vs similar alternatives. 'List all Stitch projects' (24 chars) and 'Get a specific screen' (20 chars) are below the 194-char baseline and do not explain WHEN to call them or WHAT the agent should expect. Descriptions like 'Generate a single page UI design' (31 chars) lack actionable context.
Document output schemas for all 11 tools. For each tool, specify: What fields are in the response object? What are their types? Example: sp_generate_page should return {projectId, screenId, html, framework, componentLibrary, createdAt}. Use TypeScript interfaces or JSON Schema to make this explicit.
Expand tool descriptions to 100-200 characters and answer three questions: (1) WHAT does this tool do? (2) WHEN should the agent call it (vs similar tools)? (3) WHAT will the result look like? Example: 'Generate a single page UI design based on a text description. Call this when the user requests a new page design, not when modifying an existing screen. Returns the generated HTML, framework code, and design system mappings.'
Implement error categorization and recovery guidance. For write operations (generate_page, create_design_system, etc.), catch exceptions and return structured errors: {status: 'error', code: 'INVALID_PROMPT', message: 'Prompt too short (min 5 chars). Try adding more detail.', retryable: true}. Include error codes (INVALID_PROMPT, PROJECT_NOT_FOUND, DESIGN_SYSTEM_MISMATCH) so agents can distinguish recoverable from fatal failures.
Add dry-run or confirmation modes for destructive/write operations. For sp_generate_page, sp_apply_design_system, and sp_convert_html, support a 'dryRun: true' parameter that returns a preview without persisting changes. This prevents agents from accidentally overwriting screens.
Expand enum value descriptions. When framework accepts [html, react, vue, svelte], document each choice in the parameter description: 'Target framework. html: vanilla HTML/CSS/JS; react: React with optional Next.js; vue: Vue 3; svelte: Svelte 5. Default: html if no design system, react if design system is applied.'
Spec posture evidence
Inferred effective spec: <=2025-11-25.
Relies on Logging (deprecated) - log to stderr or use OpenTelemetry
Score history
Overall score trend
First recorded score · v2 rubric
50/100
Scored
Grade
Overall
Spec posture
Rubric
2026-09-23
D
50
<=2025-11-25
v2
auth
50/100
Get a specific screen from a Stitch project
sp_list_projectsread onlyauth50/100
List all Stitch projects
sp_list_screensread onlyauth50/100
List all screens in a Stitch project
sp_responsive_adaptwrite50/100
Make HTML responsive by injecting Tailwind breakpoints and container queries
No error handling, recovery guidance, or categorization visible. Write operations (sp_generate_page, sp_generate_flow, sp_create_design_system, sp_apply_design_system, sp_a11y, sp_responsive_adapt, sp_convert_html) lack confirmation steps or dry-run modes. Agents have no guidance on what to do if a generation fails.
Parameter descriptions for framework, componentLibrary, and deviceType enum choices are minimal. 'Output framework' and 'Component library to map to' lack guidance on when to use react vs vue vs svelte, or when shadcn vs radix vs mui is appropriate.
No pagination or result-limiting logic visible. Tools like sp_list_projects and sp_list_screens do not declare max result counts, and there is no visible implementation of limit/offset or cursor-based pagination. Large result sets could blow context windows.
No permission checks, audit trails, or scope declarations visible in source. Write operations (create_design_system, apply_design_system, generate_page, generate_flow) do not verify that the calling agent or user has authority to perform these actions. No logging of who did what when.
Tools returning HTML/design system data do not document field structure. sp_get_screen returns HTML 'source' but no schema explains what fields are in the response object. sp_generate_page returns generated UI but no documentation of return shape.
Implement pagination for list operations. Add limit (default 20, max 100) and offset parameters to sp_list_projects and sp_list_screens. Return {items: [...], total: N, nextOffset: M} so agents can fetch large lists incrementally without blowing context.
Add permission checks and audit logging. Before executing sp_create_design_system or sp_generate_page, verify the calling user has 'design:write' scope. Log all write operations: {timestamp, userId, tool, projectId, action, result}. Return 403 Forbidden with a clear message if the user lacks permission.
Validate input constraints explicitly in tool handlers. For sp_generate_page's prompt parameter (minLength: 5), catch short inputs and return: {status: 'error', code: 'INVALID_PROMPT', message: 'Prompt must be at least 5 characters. You provided: "abc". Please expand the description.', suggestions: ['Describe the page layout and purpose', 'Mention the target audience or use case']}
Document parameter relationships and dependencies. For sp_generate_page, explain: 'If designSystemId is provided, accessibility and responsive are automatically enabled. If componentLibrary is none, framework must be html. If framework is not react, appRouter is ignored.' Add these as notes in parameter descriptions.
Return chaining IDs in responses. When sp_create_design_system returns a new design system, include {designSystemId, projectId, name, createdAt}. When sp_generate_page returns a screen, include {projectId, screenId, html, framework}. This allows agents to immediately call downstream tools without extra lookups.