Vektos gives AI coding agents the same tasks with levels of help from nothing to a design-system MCP with a checker, and measures whether their code follows a governed design system, and what each run costs. These are the banking tasks.
643 runs in 120 batches, 6 models, $5.55 spent on runs.
With ds-mcp-check (the MCP and a checker), design-system violations fell from 102.8 to 0.1 per run, across the 6 models that ran both, for a median 1.3× the cost per run.
Each row gives the agent more help than the one above it. Every cell is the mean design-system violations per run; open one for its runs.
| Context | Mean | claude-haiku-5.5 | deepseek-v4.1-flash-1017 | kimi-k3.1-pro | nemotron-3.5-lightning-free | gpt-6-luna | beam-free |
|---|---|---|---|---|---|---|---|
| none File tools only | 102.8 | 36.5 | 94.6 |
Which context gives compliant code at the lowest cost? The best place is the bottom left.
| 294.3 |
| 10.4 |
| 141.7 |
| 39.3 |
| storybook-mcp Storybook's own MCP | 8.0 | 4.3 | 0.4 | 0.4 | 6.1 | 36.2 | 0.7 |
| ds-mcp Our design-system MCP | 0.9 | 0.9 | 0.2 | 0 | 1.6 | 2.1 | 0.3 |
| ds-mcp-check The MCP and a checker | 0.1 | 0 | 0 | 0 | 0.4 | 0 | 0 |
| ds-mcp-check-full A checker with compiler and linter | 0.1 | 0.2 | 0 | 0 | 0.5 | 0 | 0 |
| genui-spec A view from the catalog, six tasks | 0 | – | – | – | 0 | – | 0 |