frameevidence

When an agent implements a Figma design, it often guesses field names, style tokens, and component hierarchy, producing code that drifts from the source design. FrameEvidence closes that gap by handing the agent a bounded, read-only slice of design evidence before a single line of code is written: layout structure, typography variables, and component references pulled through the Figma REST API at a limited depth and node count.

Two read-only tools do the work. figma_inspect returns an implementation-focused node tree rather than dumping the full Figma JSON schema into the model context. figma_render yields a temporary PNG, JPG, SVG, or PDF render URL for a chosen node. The entire pipeline is strictly read-only: the personal access token is sourced exclusively from a host environment variable and never appears in model arguments or output; requests target api.figma.com only and reject all redirects; raw responses are byte-capped, node trees are bounded in depth and count, and an in-memory cache reduces rate-limit pressure.

FrameEvidence fits developers running Figma-to-code workflows on DeepSeek Harness or similar agent frameworks; teams that need to inject minimal usable design context without introducing OAuth, multi-user credential stores, or a maintained server; and security owners who want a clear, auditable boundary around agent access to design files.