Secret MCP: Evidence-Bounded and Context-Isolated Design Specification Generation from Web Screenshots

Abstract

Screenshot-to-code systems optimize for rendered implementations, but screenshots omit document structure, interaction logic, responsive rules, and provenance needed to distinguish observation from guesswork. Multi-reference prompts also risk contaminating one reference with evidence or inferences from another. We present Secret MCP, an open-source local system that produces one auditable design specification per public web reference. It separates retrieval, evidence preparation, model invocation, storage, and inspection. Long captures are resized and tiled with overlap; evidence records preserve prepared- and source-space coordinates and a measured color palette. A 19-section contract covers page inventories, navigation geometry, responsive matrices, components, accessibility, acceptance criteria, and explicit labels for measured, observed, inferred, and unknown claims. References are processed sequentially through a sampler interface. The evaluated MCP adapter sends one sampling/createMessage request per reference with includeContext set to none; a fresh-process adapter provides a stronger boundary. We evaluate commit c130c9c at two levels. A live retrieval and fixture-model integration run selected two references after excluding a third, prepared nine evidence images, issued two sampling requests, and produced two documents with zero cross-reference identifier occurrences. A static audit of three externally generated design indexes found all 19 required sections in every document, one unique reference identifier per document, eight page specifications, 1,943 pixel-valued measurements, and 319 color literals. These tests establish orchestration invariants and syntactic contract compliance, not semantic or visual reconstruction accuracy. The sampler abstraction preserves these boundaries across direct model APIs and transports despite MCP sampling's 2026 deprecation.