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.

Feedback wanted: say where it breaks in the launch feedback thread, see what's next in the roadmap, or pick up a good first issue. If onsystem is useful to you, a star helps other people find it.

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.}

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 <button>, variant="danger" on a Button that only has destructive, and <Card.Header> 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.

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.

Quickstart

Run it from the app's folder or from the monorepo root, with Node.js 20.19 or later:

npx -y onsystem inspect   # what it found; --explain says why
npx -y onsystem check .   # the same check the agent and CI run
npx -y onsystem init      # write the config, and offer a baseline

At a monorepo root, inspect lists every app and design-system package, and check checks each file by its own project's design system. In a design-system package, init sets it up for every app that depends on it, with no config in the apps: as a workspace package, or installed from npm when it publishes its component sources (init checks its files list). From a package that publishes only compiled code, installed apps get the token rules but no component, prop, variant or native-element checks.

In a new create-next-app, the template page's hardcoded colors already make the first check . fail; init offers to record them as a baseline, so only new findings fail. Folders installed from a shadcn registry, such as components/ai-elements, are someone else's code: Code from shadcn registries says how to check them or leave them out.

Claude Code. The plugin brings the hook, the MCP server and a skill:

/plugin marketplace add dgesteves/onsystem
/plugin install onsystem@dgesteves

After each Write or Edit of a .tsx or .jsx file, the hook runs onsystem check on it. Errors go back to Claude as a blocking error with the fix, so Claude corrects the file before it moves on; warnings go as a note. It stays quiet in projects without a design system. Codex installs the same plugin, and the hook checks the files its apply_patch changes. What it runs and touches.

Cursor or VS Code. , or commit the config. The agent gets the tools and runs check_ui on what it writes.

Other agents. npx skills add dgesteves/onsystem adds the skill to any agent with Agent Skills, and there are plugins for Codex, GitHub Copilot CLI, VS Code and Cursor.

Any MCP client. Run npx -y onsystem as a stdio server. There are configs for Claude Desktop, Codex CLI, GitHub Copilot CLI, Windsurf, JetBrains IDEs, Zed, Gemini CLI, Antigravity and Grok Build.

CI. Install it as a dev dependency, accept the findings the codebase already has once, and fail pull requests only on new ones:

npm install --save-dev onsystem
npx onsystem check . --update-baseline   # writes onsystem.baseline.json: commit it

- run: npx onsystem check . --format github --require-design-system

Findings show up as annotations on the pull request, and --require-design-system fails the job when the design system is no longer found. CI and baselines has the whole workflow.

GitHub Action. The action runs the same check, honours the baseline and annotates only the lines a pull request changed. It goes after the checkout and the install, so it can read your design system:

steps:
  - uses: actions/checkout@v7
  - uses: actions/setup-node@v7
    with:
      node-version: 24
      cache: npm
  - run: npm ci
  - uses: dgesteves/onsystem@v0

Make it a required check on agent pull requests has the whole workflow and its inputs.

ESLint. The same rules for ESLint 8.57, 9 and 10: add onsystem.configs.recommended from onsystem/eslint to eslint.config.mjs (ESLint plugin, with @shadcn/lint).

Works with @shadcn/lint

@shadcn/lint checks the Tailwind classes written against a component and the theme (restyling, raw colors, arbitrary and unknown classes) as ESLint or Oxlint rules on Tailwind v4. onsystem checks that the components, props and variant values exist and flags native elements and unnamed icon buttons, on Tailwind v3 or v4. Its MCP server shows the agent your components and tokens before it writes, and in Claude Code its hook checks every edit. Run both. Benchmark v2 also scores every output with @shadcn/lint: see the independent scorer's table.

Does it help?

Benchmark v2 gave 6 models from 3 providers (Anthropic, OpenAI and Google) 20 single-file UI tasks in 5 open-source repositories, without onsystem, with its skill, MCP server or hook, and with all three: 1,200 runs, 240 per arm. Counted by onsystem 0.5.0's own check, design-system errors per run went from 0.85 without onsystem to 0.00 with all three, with no measurable change in the tsc pass rate (93% → 92%).

That count is onsystem's own, and the MCP server and the hook show the agent the same errors to fix. In tasks that ask for a green status label in a design system with no success token, agents with onsystem often used chart-* tokens, which it counts as clean. The blind labels by an AI reviewer, 12 files per arm, are too few to confirm the count either way, and every violation they found with the hook or all three is a palette color in Dub, whose tokens 0.5.0 did not read. The benchmark page has every table, and bench/v2 the method, the per-run data and every generated file.

Docs

On the website, and as Markdown in docs/:

• Quickstart, Claude Code plugin, CI and baselines and Set up your agent, with the config for each client
• Configuration, Rules, MCP tools and Troubleshooting
• How it works, FAQ and Migrating from design-system-mcp

Contributing, security and the changelog are in the repository.

License

MIT © Diogo Esteves