dgesteves/onsystem

**Keeps coding agents on your design system:** catches palette colors, native elements and APIs your system lacks, at every edit and pull request, with the fix. Set it up once in your design-system package; every app in your monorepo gets it. Runs locally and works alongside @shadcn/lint.
Built for design-system and platform teams with a React design-system package; a shadcn/ui app is the zero-config demo. Formerly @dgesteves/design-system-mcp (what changed).
Using the GitHub Action? uses: dgesteves/onsystem@v0 checks the lines a pull request changed: see CI and baselines.
**Also works with**
**Website and docs:** onsystem.vercel.app, with a playground that checks your own code.
https://github.com/user-attachments/assets/900cc1fa-387c-4e76-a524-a818521dd1a0
On the website , an agent drafts a settings card, check_ui finds 11 problems with a fix each, and the agent applies them. Then a real Claude Code session with the plugin: Claude looks the components up before writing, and check_ui passes. Last, the benchmark.
On the website , an agent drafts a settings card, check_ui finds 11 problems with a fix each, and the agent applies them. Then a real Claude Code session with the plugin: Claude looks the components up before writing, and check_ui passes. Last, the benchmark.
-->
The problem
Coding agents write UI from their training data, not from your design system. Ask for a settings card and you get bg-[#ef4444], p-[13px], a native , variant="danger" on a Button that only has destructive, and where the system exports CardHeader. On the demo draft above, tsc reports 3 of the 11 problems, and review catches the rest when someone has the time. Describing the design system to the agent helps; what holds is checking what it actually wrote, at the edit and at the pull request.
What it checks
onsystem reads your components, cva and tv variants, parts and tokens from source with the TypeScript compiler, and checks .tsx and .jsx against them. Every finding has a rule id, a location and the fix. Tokens come from your stylesheets, DTCG 2025.10 files or a DESIGN.md, and onsystem export writes them out as either.
| Rule | Default | Catches |
| -------------------------------- | ------- | ------------------------------------------------------------------------------- |
| no-unknown-component | error | Components and parts that don't exist ( , typos) |
| no-unknown-prop | error | Props a component doesn't take (tone, isDisabled) |
| no-unknown-variant | error | Values outside a variant (variant="danger") |
| prefer-design-system-component | error | Native elements the design system has a component for ( → ) |
| no-hardcoded-color | error | Hex, rgb, hsl, oklch and Tailwind palette colors (bg-gray-100 → bg-muted) |
| icon-button-accessible-name | error | Icon-only buttons without an accessible name |
| no-hardcoded-spacing | warn | Arbitrary padding, margin and gap (p-[13px] → p-3) |
| no-hardcoded-radius | warn | Arbitrary radius (rounded-[7px] → rounded-sm) |
**How often it is wrong.** The corpus job runs check on 12 public repositories pinned by commit (Documenso, Dub, Cal.com, midday, Flowise and others: 24 runs, 14,539 distinct findings), without their dependencies installed (the two MUI apps get MUI). A seeded random sample of 80 of those findings, stratified by rule, was labelled from the code alone: blind labels by an AI reviewer (Claude Opus 5.5), not by a person. 57 are right, 14 debatable and 9 false positives: 11.2% (95% Wilson interval 6.0-20.0%), about the same when weighted by each rule's share of all findings (11.0%). Most come from no-hardcoded-color (7 of 65 sampled). Counting values a project's own design system already uses as debatable, 6 of 80 are false positives (7.5%, 3.5-15.4%). The sample has each label with its reason, the labeller's notes and the script behind these numbers. Shipped code rarely invents components, props or variants, and without dependencies no-unknown-prop skips components whose props do not resolve: those three rules have 56 findings in the corpus, 55 of them in Flowise, an MUI app, and 3 in the sample, all right but too few for a rate of their own. The corpus's older labels, made by an AI reviewer who could see the finding and so not blind, fail the job when one of their true positives stops being reported or a fixed false positive comes back. Code installed from a shadcn registry, such as the chatbot's components/ai-elements, has no labels yet (Code from shadcn registries). The rules reference shows each rule on an example and how to suppress a finding.