Decantr MCP Server

Decantr

AI Frontend Governance for production codebases touched by AI agents.

Decantr is the governance layer between product intent and AI-edited frontend code. It gives coding assistants three things they don't have on their own: typed contracts, route-scoped context, and machine-actionable evidence, so UI changes stay coherent instead of drifting prompt by prompt. The model still writes the code; Decantr defines the contract, context, and verification loop around it.

AI generates the interface. Decantr proves, with local evidence, whether the outcome stayed aligned.

Pick your path

---

Greenfield Content Contracts

1. Scaffold from a blueprint

npx @decantr/cli new my-app --blueprint=esports-hq
cd my-app

A blueprint is official corpus content: theme, sections, pages, layouts, voice, and personality that compile into Decantr contracts and execution packs. It is reference material for governance, not a public marketplace or runtime takeover. Try esports-hq, agent-studio, or carbon-ai-portal, or run decantr list blueprints --blueprint-set featured to browse the curated set.

Adapter availability. Greenfield defaults to contract-only governance. The legacy @decantr/css runnable adapters (react-vite, next-app, vanilla-vite, vue-vite, sveltekit, angular, and solid-vite) remain available only when explicitly requested with --adoption=decantr-css. Other targets initialize through generic-web contract-only mode.

2. What just got generated

my-app/
├── decantr.essence.json     # the durable contract: theme, sections, routes, features
├── DECANTR.md               # methodology primer your AI assistant reads first
├── .decantr/context/
│   ├── scaffold.md          # full app overview: topology, voice, personality
│   └── section-*.md         # per-section spec: shell, patterns, spacing
└── src/styles/
    ├── tokens.css           # CSS variables from the theme
    ├── treatments.css       # shared visual treatment classes
    └── decorators.css       # theme-specific decorator classes

Decantr produces the contract. Your AI assistant produces the implementation against it.

3. Hand it to your AI assistant

Open the project in Claude Code, Cursor, Windsurf, or any AI-aware editor. For Cursor, run decantr connect cursor once from the opened workspace; in monorepos use decantr connect cursor --project apps/web. It writes .cursor/mcp.json and .cursor/rules/decantr.mdc, preserving existing MCP servers, so Cursor Agent knows to call Decantr task context before route edits. Other assistants can read DECANTR.md first for the methodology, then load section context files on demand. For route-level work, run decantr task <route> "<intent>" before editing so the assistant gets the active authority block, local law, evidence, and verify command. The split keeps the assistant focused on the right scope at the right time.

4. Make your first change and verify

# Edit decantr.essence.json — add a section, swap the theme, etc.
decantr refresh   # regenerate context files from the updated essence
decantr graph     # generate the typed Contract graph, source index, and contract capsule
decantr verify    # run the canonical local reliability gate
decantr verify --evidence
decantr studio    # open the local-only health dashboard

refresh keeps the generated context files in sync with the essence. graph writes .decantr/graph/graph.snapshot.json, .decantr/graph/snapshots/<snapshot-id>.json, graph.manifest.json, graph.diff.json, and contract-capsule.json for typed agent context and replayable local history. verify is the day-to-day reliability gate over Project Health, Brownfield checks, health baselines, and optional evidence; health findings include stable diagnostic codes, typed repair IDs, component reuse drift such as COMP001 when an agent reimplements a project-owned primitive, behavior-obligation drift such as A11Y010 when an accepted confirmation dialog loses its accessible name, and accepted style bridge drift such as TOKEN010 when arbitrary Tailwind, inline style, or stylesheet values bypass project-owned style authority. When a graph snapshot exists, findings also include graph anchors back to the contract node they came from. --evidence writes the privacy-redacted Evidence Bundle that AI agents and CI can use for repair context, including those codes, repair IDs, and graph anchors when available. decantr studio gives the same signal in a small localhost dashboard, and decantr studio --workspace expands that view across many Decantr projects in a monorepo.

Starting from a different point? See the full workflow model.

---

The model

Decantr separates design governance into two layers:

• DNA — durable visual and system axioms: theme, spacing, motion, accessibility, personality.
• Blueprint — product topology: sections, page routes, shells, layouts, features.

That split matters because not every change should be treated the same way. A theme swap or accessibility regression is different from adding an auxiliary section or reshaping a route map. Decantr lets governance be strict where it should be strict (DNA, errors by default) and flexible where it should be flexible (Blueprint, warnings only).

Canonical shapes live in the published schemas; the command-level flow is in the workflow model.

Surfaces

Decantr 3.7 introduced the shared Brownfield discovery substrate that Decantr 3.8 now uses across verifier, CLI, MCP, and app-scoped workflows. scan, setup, workspace list, adopt, doctor, task, verify, ci, and MCP project/task/evidence reads agree on selected app scope, workspace package manager, framework, language, route signals, taskable routes, component inventory confidence, styling authority, assistant rules, and limitations. This is especially important in mixed monorepos: a React/Vite app beside an Angular app should scan as React/TypeScript/pnpm when --project apps/web targets the React app, not as the repository root or sibling framework. scan --json emits scan-report.v2 by default; scan-report.v1 remains published as a historical schema reference.

Security review starts with the installed npm surface, not the whole monorepo. The package permission matrix documents filesystem, network, process, telemetry, hosted-upload, and MCP write-tool behavior for every public package: docs/reference/security-permissions.md.

Packages

Full release/support status lives in docs/reference/package-support-matrix.md.

Repo layout

Showcase capsule architecture is documented in docs/reference/showcase-host.md.

Development

Requires Node.js >=20.19.0 and pnpm >=9.

pnpm install
pnpm build
pnpm test
pnpm lint
pnpm audit:public-api
pnpm audit:package-permissions
pnpm audit:content-package
pnpm showcase:verify:shortlist
pnpm benchmark:realworld-corpus -- --config scripts/realworld-corpus.first-mile.json --cli packages/cli/dist/index.js --out /tmp/decantr-realworld-corpus-3.8
pnpm release:verify
pnpm release:closeout

Release stewardship

Release work is script-led. Use pnpm release:commands for publish handoff, pnpm release:verify for public npm verification, and pnpm release:closeout before calling a release done. Closeout verifies npm version/dist-tag parity plus git tag and release-note parity so versioned docs cannot drift away from origin tags. After closeout, pnpm release:announce -- --version X.Y.Z --send can dispatch the changelog to community-ops for Discord posting. See docs/runbooks/release-stewardship.md.

More CLI

Intent and discovery:

decantr setup
decantr magic "AI chatbot with a bold terminal-inspired workspace"
decantr search dashboard
decantr list blueprints --blueprint-set certified
decantr list blueprints --blueprint-set labs
decantr suggest leaderboard
decantr suggest "recipe feed with infinite scroll" --route /feed --from-code
decantr suggest --from-code --file app/page.tsx --project apps/web
decantr suggest "standardize buttons" --project apps/web

Source-code suggestions rank official content patterns and accepted .decantr/local-patterns.json against the selected app's file contents. From an app root, decantr suggest "button" --from-code --file src/App.tsx surfaces project-owned law without extra flags; from a monorepo root, add --project apps/web.

Brownfield adoption:

decantr scan
decantr adopt --yes
decantr studio
decantr doctor
decantr codify --from-audit --style-bridge
decantr codify --map-pattern hero
decantr codify --accept
decantr graph
decantr graph --file src/app/page.tsx --impact --json
decantr resolve
decantr connect cursor
decantr task /feed "add saved recipe actions"
decantr verify --brownfield --local-patterns
decantr verify --since-baseline

In a monorepo, install Decantr at the workspace root, but attach it to an app:

decantr setup
decantr workspace list
decantr scan --project apps/web
decantr adopt --project apps/web --yes
cd apps/web && decantr studio
decantr doctor --project apps/web
decantr codify --from-audit --style-bridge --project apps/web
decantr codify --map-pattern hero --project apps/web
decantr graph --project apps/web
decantr graph --project apps/web --file src/app/page.tsx --impact --json
decantr resolve --project apps/web
decantr connect cursor --project apps/web
decantr task /feed "add saved recipe actions" --project apps/web
decantr verify --brownfield --local-patterns --project apps/web
decantr ci init --project apps/web
decantr ci --project apps/web
decantr add page app/settings --route /settings --project apps/web
decantr health --project apps/web --prompt <finding-id>

For observed Brownfield apps, app/settings can resolve to the single primary section, such as observed-primary, so route additions remain ergonomic after attach.

If the app is already running and you want local screenshots, add visual evidence later:

decantr verify --project apps/web --base-url http://localhost:3000 --evidence

analyze now writes Brownfield intelligence artifacts in addition to the proposal: .decantr/brownfield-intelligence.json, .decantr/theme-inventory.json, and .decantr/enrichment-backlog.md. Essence V4 stays unchanged; multi-theme apps are observed in the inventory and kept as local task context until the team explicitly promotes that model. When local law, a style bridge, or Decantr CSS is accepted, Decantr calls that out as Hybrid instead of pretending every existing app is still only contract-only.

scan is the safe first look: it is read-only and tells you whether Decantr sees a plausible UI app before anything is written. adopt orchestrates the primitive Brownfield flow (analyze, proposal acceptance, optional content pack hydration when online, first typed Contract graph baseline, Project Health, optional browser evidence, optional CI) so users do not need to memorize the internal command chain. In contract-only/offline adoption, content packs can be deferred; doctor, health, and refresh --check treat missing packs as optional context unless a present pack-manifest.json references missing files. studio gives a local visual view of what Decantr found, while doctor explains project/workspace state, generated artifacts, typed graph freshness, CI wiring, and the active adoption lane: Brownfield contract-only, Hybrid local law, Hybrid style bridge, Hybrid Decantr CSS, Hybrid composition, or Greenfield. codify --from-audit proposes project-owned local law in .decantr/local-patterns.proposal.json and .decantr/rules.proposal.json, with source evidence, confidence tiers, likely variants for button/card/form/theme families, and optional behavior_obligations for observed forms or destructive confirmation dialogs; codify --style-bridge proposes .decantr/style-bridge.proposal.json, mapping Decantr intent to project-owned tokens/classes without installing Decantr CSS; codify --map-pattern <slug> maps official content guidance into an advisory local-law proposal without changing source. After review, codify --accept promotes whichever proposals exist so task, suggest, ci, Project Health, and MCP task context treat them as Hybrid authority, while official corpus patterns remain optional guidance until mapped into project-owned law. When typed graph artifacts exist, task and MCP task context rank route graph nodes with the task text so matching components, local rules, behavior obligations, findings, evidence, and source artifacts rise before the edit. The primitive commands remain available for advanced and scripted workflows.

Content and verification:

decantr content summary       --namespace @official --json
decantr content compile-packs decantr.essence.json --write-context
decantr content compile-packs apps/web/decantr.essence.json --write-context
decantr audit src/pages/Home.tsx
decantr audit
decantr showcase verification  --json

When an essence path is provided, --write-context writes the pack bundle beside that essence file. In monorepos, that means app packs land under apps/web/.decantr/context, not the repository root.

Project Health remediation copy follows the same rule: from a monorepo root, pack fixes point at apps/web/decantr.essence.json, prompt commands keep --project apps/web, task/read paths include apps/web/..., and CI prompts point at decantr ci --project apps/web --fail-on error. ci prints accepted local-rule findings with file and line evidence, even when warnings are not configured to block. task also prints an authority block so assistants know whether the route is Brownfield contract-only, Hybrid local law, Hybrid style bridge, or Decantr CSS, and it warns before mixing runtimes or adding Decantr CSS to a non-Decantr-CSS app. Advanced app-scoped primitives such as health, status, add, remove, theme, export, suggest, and magic should target the selected app rather than the workspace root. setup becomes a post-adoption dashboard once any app is attached, so it recommends doctor, task, verify, and ci init instead of telling users to adopt the same app again.

Project Health, CI, and diagnosis:

decantr doctor
decantr doctor --project apps/web
decantr verify
decantr verify --base-url http://localhost:3000 --evidence
decantr verify --since-baseline
decantr ci --project apps/web
decantr ci --workspace --changed --since origin/main
decantr ci init --project apps/web
decantr ci init --workspace
decantr health --format markdown
decantr health --prompt <finding-id>
decantr health --evidence --output .decantr/evidence/latest.json
decantr health --browser --base-url http://localhost:3000 --evidence
decantr health --design-tokens .decantr/design/figma-tokens.json
decantr verify --workspace --changed --since origin/main
decantr export --to figma-tokens
decantr health --json --output decantr-health.json
decantr studio --port 4319 --host 127.0.0.1
decantr studio --workspace
decantr studio --report decantr-health.json

In contract-only Brownfield projects, Decantr may not own src/styles/tokens.css; export --to figma-tokens is for Decantr CSS token exports, while project-native token exports remain the app's source of truth unless you adopt a style bridge.

Caller-configured private telemetry:

decantr telemetry status
decantr telemetry explain
DECANTR_TELEMETRY_ENDPOINT=https://telemetry.example/v1/events decantr init --telemetry
decantr telemetry link --api-url https://telemetry.example/v1 --api-key <key>

Decantr does not operate a hosted telemetry sink. Setting telemetry: true records a local preference, but the CLI sends nothing and creates no opaque telemetry identifiers unless DECANTR_TELEMETRY_ENDPOINT points to a caller-controlled sink. Private deployments can additionally configure identity linking with an explicit --api-url; the Decantr content API does not provide that route. When configured, telemetry stays product-level: command names, aggregate lifecycle counts, content sources, Project Health scores/counts, and adoption workflow signals. It never includes source code, prompts, raw paths, emails, private package slugs, or health report contents.

Official content supply chain:

decantr content check
decantr content check --ci --fail-on error
decantr content-health
decantr content-health --markdown --output content-health.md
decantr content-health --ci --fail-on error
decantr content check --prompt <finding-id>

FAQ

How does Decantr keep AI or agent work tied back to the original spec?

You cannot truly force an LLM to follow a spec. Decantr gives the assistant a durable contract, scoped context, and drift checks so implementation stays tied to the intended product shape.

decantr refresh
decantr verify
decantr audit
decantr health

If the product direction changes, update Decantr deliberately, then regenerate and check:

decantr add page dashboard/reports
decantr refresh
decantr verify

How do I open Decantr Studio?

Run Studio from the app root, where decantr.essence.json lives:

decantr studio

If the default port is busy:

decantr studio --host 127.0.0.1 --port 4320

Studio is local-only and shows the same v2 Project Health signal as decantr health. Its Control Room keeps triage focused: inspect loop state, next action, authority lane, blocking findings, evidence tier, graph impact, and copyable commands. Workspace mode is available with decantr studio --workspace when a repo contains many Decantr projects.

decantr health --prompt <finding-id> prints a focused repair prompt for the assistant doing the implementation. It does not edit files by itself; use Studio's copy buttons or paste the printed prompt into your AI coding workflow, then rerun Project Health.

To view a CI-produced report artifact without scanning a checkout:

decantr health --json --output decantr-health.json
decantr studio --report decantr-health.json

Studio is for interactive triage; decantr ci --fail-on error is the pull-request gate.

See the full FAQ for common setup, brownfield, Studio, migration, CI, and drift questions.

Links

• Discord — discord.gg/WeDpBd4xFU
• FAQ — docs/faq.md
• Monorepos — docs/guides/monorepos.md
• Published schemas — decantr.ai/schemas
• Project Health — docs/reference/project-health.md
• Command surface — docs/reference/command-surface.md
• Content Health — docs/reference/content-health.md
• Workflow model — docs/reference/workflow-model.md
• Content API reference — docs/reference/registry-public-api.md
• Package support matrix — docs/reference/package-support-matrix.md

Contributing

Contributions are welcome. The most useful repo context lives in docs/ — especially the architecture notes, audits, runbooks, and package support matrix.

Use Discord for quick setup help, showcase feedback, and live community discussion. Keep bugs, feature requests, and durable decisions in GitHub issues, PRs, or docs.

License

MIT. The source repositories are MIT licensed; hosted services such as the content API may publish separate service terms without changing the source-code license.