What Saer checks

Tokens across themes and modes, exact drift, near-matches, dead and conflicting declarations.

• Reads Tailwind v3 configs, Tailwind v4 @theme blocks and CSS custom properties, including dark mode and other theme contexts.
• Follows every var() chain to a real value in every context.
• Exact drift: a raw value identical to a declared token. Deterministic.
• Near-matches: raw values perceptually close to a token, with the distance, always kept apart from exact drift.
• Dead declarations and conflicting definitions.

Families, sizes, weights, line heights and tracking: the type scale you really use, and what sits off it.

• Declared scales from Tailwind v3 and v4, utility classes and CSS custom properties.
• Usage from classes, CSS, inline styles and CSS-in-JS.
• Exact drift, dead tokens and, labelled experimental, off-scale values with the nearest step.

Spacing, sizing, radius, shadows, z-index, breakpoints, borders, opacity and motion.

• The scales hiding in your styles, declared or in use.
• Exact drift and, labelled experimental, off-scale values, with the nearest step only where the distance means something.

Inventory, use sites, duplicates, variant sprawl, one-off overrides and unused components.

• Every React component and every JSX use resolved to its definition through imports, aliases, barrels and workspace packages.
• The same component name defined in more than one file, with the evidence.
• Variant values passed but not declared, or declared and never used.
• Use sites that override styling the component already sets.

Static checks: contrast per light and dark context, focus styles, tap targets and reduced motion.

• Contrast of the text and background pairs written together in the source, judged in every colour context.
• Focusable elements that remove the outline and show nothing else.
• Pointer targets written below 24 by 24 px.
• Animations with no reduced-motion fallback.
• Static checks only: not a statement of WCAG conformance.

Point Saer at a URL for a read of a live site's compiled CSS. Polite: robots.txt respected.

• Reads public HTML and stylesheets, no JavaScript, no login, one request per second.
• Findings describe the compiled CSS that ships, which may differ from the source.

`saer mcp`, a local MCP server for Claude Code, Cursor and Codex: your tokens, components and hooks, quality warnings, a recommendation for a task, and a self-check of the agent's own changes.

• Gives a coding agent the system Saer reconstructed, so it can reuse a token, component or hook instead of inventing one.
• `saer_check_changes` checks only the lines the agent wrote, against the same rules as a scan.
• Local and read-only, like the CLI. No model at answer time: recommendations are deterministic and cite their evidence.

Explore a project in the terminal: tokens, component usage, task recommendations and change checks, with source evidence.

• Run saer, then use /scan, /overview, /tokens, /components, /recommend and /changes.
• Your project stays local. Recommendations are leads to inspect, not automatic code changes.

Connect once for automatic full-scan and CI summary history, changes between scans and revocable scoped keys.

• Source code, paths, token values and full reports stay local. Extra summary fields are refused.
• Upload keys are hashed, project-scoped and expiring; they cannot download the engine or manage an account.
• CI runner keys have a separate scope for protected headless analysis. Managed scheduling and team governance remain planned.

Compare changed lines with an explicit base commit and produce a machine-readable CI result.

• Runs on Apple Silicon Macs, with an interactive sign-in or an expiring headless runner key.
• Exit 1 for deterministic drift; incomplete checks exit 2. Review findings are advisory unless --strict is selected.
• CI runner keys permit analysis and project uploads, with expiry and revocation. Managed team policies and workload-bound identities remain planned.

Evidence-led next actions in your scan, terminal and coding-agent overview, with review-only foundation starters where no declarations were established.

• Existing systems receive an adoption plan. Incomplete coverage and unresolved evidence stay explicit; no organisational maturity score is invented.
• Starter colours, scales and optional React primitives are proposed choices. Observed colour candidates retain their source classification; semantic roles are not inferred.
• Review the local preview and approve chosen changes in your normal agent workflow. Saer never applies the proposal or uploads it, and generated suggestions are excluded from future analysis.

A local design-system reference generated from the same evidence as your scan: tokens, components and contracts, with unknowns kept explicit.

• Premium scans refresh .saer/reference.md. It is a bounded snapshot; cloud documentation and a shared registry remain planned.

A measured comparison of coding-agent output with and without Saer, on real codebases.

• Three rounds have run; the latest (Codex on three real codebases, 48 paired tasks, blind-judged) is written up. Results are published with their full method, including where Saer did not help, once reviewed.

One report across every area, with scores you can trace and the top fixes.

Managed pull-request policies, shared CI identities and continuous team drift monitoring.

A shared, versioned record of a team's system.

Adapters beyond the web.

Compare the system in code with the system in design files.