Not “pixel-perfect”.
Provably pixel-accurate.

The handoff nobody has fixed

Design-to-code is the most expensive, least reliable step in the delivery pipeline. It is done by hand, differently by every developer, and it drifts the moment the design changes.

The existing tools all produce a starting point that a human then rewrites. None produce a repeatable, updatable rail: a pipeline you re-run against an updated Figma and get a reviewable diff rather than a fresh pile of code. And not one of them publishes a substantiated accuracy benchmark. “98% pixel perfect” appears in marketing copy with no methodology, no dataset, and no definition of the metric.

CONCRA’s wedge is not generation. It is measurement.

How it works

1. Extract

   Drives Figma directly and reads every frame, every node, every variable. No plugin to install, and no modification of Figma’s signed binary — that is circumvention, it breaks on every auto-update, and it fails enterprise review on sight.

2. Normalise

   Everything becomes CIR, a layout algebra that projects onto flexbox, SwiftUI stacks, Compose, Flutter and Android without smuggling web assumptions into a native target. Constructs with no equivalent are recorded as lossy rather than quietly approximated.

3. Generate

   Vanilla HTML/CSS/JS and React today; SwiftUI, Flutter and Android XML share the same layout algebra. Bound fills emit var(--token), never a hex code — the difference between output a team adopts and output they regenerate once and abandon.

4. Verify

   Renders what it generated, diffs it against Figma’s own render of the same frame, and names the node behind every difference. Below threshold, the build fails. This is the part no competitor has.

5. Re-run

   The design changes, you re-run, and your hand-written code survives. Generated regions are fenced; everything outside them is yours. Three-way merge against a snapshot of what was last generated, so a conflict means two people really did change the same thing.

The benchmark

Intervals are a percentile bootstrap over 10,000 seeded resamples for the means, and Wilson for the pass rate. Neither can report a value the data does not support: a t-interval on a score capped at 1.0 routinely claims a mean better than identical, and the normal approximation claims certainty from a perfect small sample. Measured on a real production design file — Momentum Bakari, 10 pages and 325 screens — not on a demo built to score well. 93 screens carry a reference render and are measured below.

Every frame, worst first

Published in full rather than as an average, because an average hides the frame that would have embarrassed us.

What CONCRA does not claim

A ten-nines accuracy claim is not achievable and, more importantly, not measurable. On a 1440×900 frame that would mean fewer than 0.0001 pixels may differ; one antialiased glyph edge breaks it. There is also no oracle: Figma’s renderer and a browser’s disagree on text by construction, and neither is wrong.

• Visual line breaks are not exposed by any Figma API. Only explicit newlines are. Any wrapped text block is reconstructed, and the residual is measured rather than hidden.
• absoluteRenderBounds is null for text inside component instances — exactly where production text lives. A documented API gap, not a bug we can fix.
• Some Figma features have no target-side equivalent at all. Noise and texture effects, progressive blur, variable-width strokes, several blend modes. These rasterise, and the report says so rather than counting the difference against codegen.
• SwiftUI cannot be pixel-verified off macOS. No cross-platform renderer exists. Off macOS it runs a structural gate, and the report never implies otherwise.

Every one of those limits is in the tool’s own output, not just on this page.

Targets

Use it from your agent

CONCRA ships an MCP server, so Claude Code, Cursor or your own tooling can drive the whole pipeline. One concra_load_file spends the Figma quota once; every call after that is local and free.

claude mcp add concra -- concra mcp serve

That quota arithmetic is the point. Figma’s own MCP server is capped at a few hundred reads a day and hands back opinionated React, not data.

Pricing

The field sits at $24–$100 per seat. CONCRA sits near the top of it, because a measured guarantee is worth more than an unmeasured one.

See it against your own file

The interesting demo is not ours. Point CONCRA at a design your team is arguing about and read the report.

Book a demo