Wonder is in public alpha. The outputs are only plausible — but that's where canvas-to-code sits generally right now, not a specific problem with Wonder.

The angle it's taking is different from most. Canvas-to-code tools typically treat the canvas and the codebase as separate worlds: generate, export, paste, clean up. Wonder's MCP integrations (Claude Code, Cursor, Codex) let the design surface read your actual codebase before generating — your components, your tokens, your structure. What comes out has a better chance of fitting what's already there than something generated from a blank slate.

In practice at alpha, you're getting 60-70% of the way to something shippable. The MCP approach is the right bet, though. A tool that knows your codebase before it starts has a structural advantage over tools that treat generation as a one-way export.

Free tier: 300 credits. Pro is $20/month for 3,000 credits and unlimited MCP calls. Worth trying at that price if you want to see what this approach looks like in practice.