Findings & checks · 5 flagged
HIGH"Ignore previous instructions" command embedded in the skillSS-SKILL-INJECT-IGNORE-01 · Prompt injection · docs/plans/2026-07-16-001-security-agentic-hardening-roadmap-plan.md
The text {match} is the classic direct prompt-injection phrasing. Placed in a skill body that the agent reads as trusted instructions, it tries to make the agent abandon its prior rules and follow whatever comes next — a full system-prompt override.
- Delete the
ignore/disregard/forget … previous instructionssentence. - If this is jailbreak-research or tutorial content, move the example into a clearly fenced, non-instruction block and label it as a quoted sample.
Trace & refs
HIGH"Ignore previous instructions" command embedded in the skillSS-SKILL-INJECT-IGNORE-01 · Prompt injection · test/fixtures/extraction-corpus/hostile/expected.json×3
The text {match} is the classic direct prompt-injection phrasing. Placed in a skill body that the agent reads as trusted instructions, it tries to make the agent abandon its prior rules and follow whatever comes next — a full system-prompt override.
Show all 3 locations
- Delete the
ignore/disregard/forget … previous instructionssentence. - If this is jailbreak-research or tutorial content, move the example into a clearly fenced, non-instruction block and label it as a quoted sample.
Trace & refs
MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · specs/brandcode-mcp-limited-client-handoff-packet.md
A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.
- Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
- Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.
Trace & refs
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Use your brand guidelines with AI
`@brandsystem/mcp` turns the brand material you already have into context AI agents can use. Give it a website, PDF guide, Figma library, local files, or a Brandcode Studio brand. It produces a portable .brand/ runtime with design tokens, voice rules, provenance, and compliance checks.
Local-first. No account required. The default Core profile exposes 12 tools covering the complete adopt → use → check loop.
Start here
Install it for your agent, then ask: “How do I use my brand guidelines with AI?”
# Codex
npx @brandsystem/mcp install --client codex --write
# Claude Code, Cursor, Windsurf, or Claude Desktop
npx @brandsystem/mcp install --client claude-code --write
# Cline
npx @brandsystem/mcp install --client cline --writeinstall is a dry run unless you pass --write. Replace claude-code with cline, cursor, windsurf, or claude-desktop as needed. JSON-file targets preserve existing settings; Cline and Codex setup delegate to their official MCP commands so their current schemas remain authoritative. Agents can follow the compact installation guide without reading this entire README.
Already configured? Tell your agent:
Use my existing brand guidelines with AI. Start from this website/PDF/Figma library and show me what needs human confirmation.
What It Solves
AI tools default to category-average output because they have no brand context. Brand guidelines live in PDFs, Figma files, and people's heads — none of which AI tools can read at the moment of creation. The dominant failure mode isn't "broken output"; it's "correct but generic" — output that passes mechanical checks but reads like a competent generalist could have made it.
This MCP server is the authoring half of the "Two MCPs, One Brand" model. It extracts brand identity from live sources, compiles it into a .brand/ directory with structured governance (anti-patterns, proof-point status, voice rules, application rules) plus DTCG tokens, brand-runtime.json, and interaction-policy.json. That directory is the portable brand runtime — the artifact that travels with your brand from surface to surface.
- Claude Design reads the
.brand/directory natively when pointed at a governed repo - Claude Code, Claude Desktop, Cursor, and Windsurf connect to this server directly over stdio and load
brand-runtime.jsonat generation time - ChatGPT and other remote-first clients consume the same runtime as an uploaded artifact (
brand-report.html/brand-runtime.json) or via a remote MCP connection — see Compatibility - @brandcode/mcp (the hosted Use MCP) serves the same runtime over HTTP for teams that want authenticated live reads at
mcp.brandcode.studio/{slug}
With brand-runtime.json loaded, agent prompts collapse from 200-400 tokens of inline brand context to just the delta. First output is on-brand. No review bottleneck.
Quick Start
1. Add to your MCP config
Copy this into .mcp.json (Claude Code), .cursor/mcp.json (Cursor), or Windsurf MCP settings. Codex users can run npx @brandsystem/mcp install --client codex --write instead:
{
"mcpServers": {
"brandsystem": {
"command": "npx",
"args": ["-y", "@brandsystem/mcp"]
}
}
}For project-level configs committed to a repository, pin a version ("@brandsystem/mcp@<version>") so collaborators never auto-run an unreviewed release; npx @brandsystem/mcp install writes a pinned spec for you.
2. Create your brand system
Tell your AI tool:
Run brand_start with client_name="Acme Corp", website_url="https://acme.com", and mode="auto"
That single command extracts colors, fonts, and logo from the website, escalates to rendered or deeper multi-page extraction when the cheap pass is weak, compiles DTCG tokens, generates design-synthesis.json + DESIGN.md, and generates a portable HTML brand report -- all in under 60 seconds.
3. What you get
.brand/
brand.config.yaml ← brand name, source URLs, session state
core-identity.yaml ← colors (with roles), fonts, logo specs
tokens.json ← DTCG design tokens
brand-runtime.json ← single-file brand context for any AI agent
interaction-policy.json ← anti-patterns, voice constraints, never-say words
design-synthesis.json ← spacing, radius, shadows, component signals
DESIGN.md ← portable design brief (agent-readable)
brand-report.html ← visual report (paste into any AI chat)
assets/logo/ ← extracted logo files (SVG/PNG)Load brand-runtime.json into any sub-agent's context. First output is on-brand. No per-prompt boilerplate.
4. Use it
Run brand_write for a social-graphic about "Q3 product launch"
The AI now has your full brand context — colors, typography, logo, anti-patterns, voice rules — and generates on-brand content.
5. Go deeper (optional)
| Session | What it adds | Command |
|---|---|---|
| 1. Core Identity | Colors, fonts, logo, tokens | brand_start (done above) |
| 2. Visual Identity | Composition, anti-patterns, illustration style | brand_deepen_identity |
| 3. Messaging | Voice, tone, never-say words, brand story | brand_compile_messaging |
| 4. Content Strategy | Personas, journey stages, themes | brand_build_personas |
Each session enriches brand-runtime.json. Stop at any point — Session 1 alone is valuable.
6. Share with your team
Run brand_brandcode_connect to save on Brandcode Studio
Your brand persists on brandcode.studio. Teammates pull the same brand into their tools. One source of truth.
What It Does
Session 1: Core Identity -- Extract colors, fonts, and logo from a website or Figma file. Compile into DTCG tokens, a structured design synthesis layer, a portable DESIGN.md, and an HTML report.
Session 2: Visual Identity -- Define composition rules, pattern language, illustration style, and anti-patterns through a guided interview. Anti-patterns become enforceable compliance rules.
Session 3: Messaging -- Audit existing website voice, then define perspective, voice codex (tone, vocabulary, AI-ism detection), and brand story through a guided interview.
Session 4: Content Strategy -- Build buyer personas, journey stages, editorial themes, and a persona x stage messaging matrix.
Each session builds on the previous. Stop anywhere -- you get value immediately.
Two Ways To Use It
Local-first MCP flow -- Start from a website or Figma file, build a .brand/ directory locally, and use it immediately in chat or code tools with no account required.
Brandcode Studio-connected flow -- Connect an existing hosted brand from Brandcode Studio, pull the packaged brand into .brand/, and keep it synced over time.
Two MCPs, One Brand
The .brand runtime is the product. Two MCPs serve it:
`@brandsystem/mcp` — Build (this package). Author and compile the .brand runtime locally. Extract from websites, Figma, and PDFs. Compile governance (anti-patterns, proof-point status, voice rules, application rules) plus DTCG tokens, brand-runtime.json, and interaction-policy.json into a single .brand/ directory. Portable, versionable, ready to commit to any repo.
`@brandcode/mcp` — Use (hosted). Connect authorized MCP clients to the live Full Brand Runtime at https://mcp.brandcode.studio/{slug} with Brandcode bearer-key auth. Agents fetch the current runtime, search approved knowledge, check drafts, retrieve package-safe assets, leave append-only review feedback, and, with explicit capture scope, queue taste captures for human review — no per-tool guideline copy, no stale snapshots, no canonical mutation from the MCP. Tagline: "Your brand, live in every AI tool."
Same .brand runtime artifact. Two consumption paths. Build authors it; Use serves it.
Phase 0 for Brandcode MCP is locked in specs/brandcode-mcp-phase-0-lock.md as the original 8-tool read/append-only surface. The current hosted implementation adds capture_taste as a scoped contribute-tier tool: it requires explicit capture scope, queues a review candidate for human review, and never promotes canon.
Hosted availability: Brandcode MCP is pre-release and available to approved clients only — it is not yet publicly launched or registry-listed. Brand data on the hosted service is client-owned; feedback is append-only; agent history is scoped and redacted. Deletion and export requests are handled through your Brandcode Studio contact. Until public launch, use @brandsystem/mcp for local build/sync, and Live Mode (brand_brandcode_live) for connected reads that refresh from the hosted runtime within a short cache TTL.
Maintainers can verify a hosted deployment end-to-end with the smoke harness (npm run smoke:hosted-mcp with BRANDCODE_MCP_SMOKE_URL and scoped test keys). It verifies MCP initialize, tools/list, the locked hosted tool order, core hosted tool calls, and read-only insufficient-scope behavior. It never hardcodes keys; missing proof inputs are reported as blocked or skipped.
Claude Design integration
The .brand/ directory is engineered as a first-class input for Claude Design. Point Claude Design at a repo that contains .brand/ — governance YAML, narrative library, proof-point files, taste notes, DTCG tokens — and it grounds on the governed brand instead of improvising from uploaded assets. This is the Deploy path: author once with @brandsystem/mcp, then every Anthropic surface (Claude Design, Claude Code, Chat via compile packs) consumes the same runtime.
Tools Reference
Tool Profiles (0.10+)
By default the server registers the Core profile — 12 tools covering the complete loop: adopt → runtime/context → create → check → export, plus the Studio connector entry points and the clarify/promote path. This keeps first-tool selection sharp for agents.
The full profile registers the entire authoring system (all tools below). Opt in via env or args:
{
"mcpServers": {
"brandsystem": {
"command": "npx",
"args": ["-y", "@brandsystem/mcp", "--profile=full"]
}
}
}Or set BRANDSYSTEM_PROFILE=full. Core tools: brand_start, brand_status, brand_runtime, brand_context, brand_check, brand_preflight, brand_report, brand_export, brand_clarify, brand_compile, brand_brandcode_auth, brand_brandcode_connect.
Entry Points
| Tool | What it does |
|---|---|
brand_start | Begin here. Creates a brand system from a website URL in under 60 seconds. Use mode='auto' for one-call setup with rendered and deep-site fallback on weak JS-rendered sites. |
brand_status | Check progress, get next steps, or see a getting-started guide if no brand exists yet. |
Session 1: Core Identity
| Tool | What it does |
|---|---|
brand_extract_web | Extract logo (SVG/PNG), colors, and fonts from any website URL. |
brand_extract_visual | Screenshot the rendered page in headless Chrome and extract computed colors, fonts, and visual context from JS-heavy sites. |
brand_extract_site | Discover representative pages, render them across desktop and mobile, capture screenshots, sample multiple components, and persist extraction-evidence.json. |
brand_generate_designmd | Generate design-synthesis.json and DESIGN.md from extracted evidence or the current brand state. |
brand_extract_figma | Extract from Figma design files (higher accuracy). Two-phase: plan then ingest. |
brand_set_logo | Add/replace logo via SVG markup, URL, or data URI. |
brand_compile | Generate DTCG design tokens, brand runtime contract, and interaction policy from extracted data. |
brand_clarify | Resolve ambiguous brand values interactively (color roles, font confirmations). |
brand_audit | Validate .brand/ directory for completeness and correctness. |
brand_report | Generate portable HTML brand report. Upload to any AI chat as instant guidelines. |
brand_init | Low-level directory scaffolding. Prefer brand_start instead. |
Session 2: Visual Identity
| Tool | What it does |
|---|---|
brand_deepen_identity | Define composition rules, patterns, illustration style, and anti-patterns (6 interview sections). |
brand_ingest_assets | Scan and catalog brand assets with MANIFEST.yaml metadata. |
brand_preflight | Check HTML/CSS against brand rules -- catches off-brand colors, wrong fonts, anti-pattern violations. |
Session 3: Messaging
| Tool | What it does |
|---|---|
brand_extract_messaging | Audit existing website voice -- fingerprint, vocabulary, claims, AI-isms, gaps. |
brand_compile_messaging | Define perspective, voice codex (tone, vocabulary, AI-ism detection), and brand story. |
Session 4: Content Strategy
| Tool | What it does |
|---|---|
brand_build_personas | Build buyer personas through a 7-question guided interview. |
brand_build_journey | Define buyer journey stages (ships with 4 proven defaults). |
brand_build_themes | Define editorial content themes balanced across awareness, engagement, and conversion. |
brand_build_matrix | Generate messaging variants for every persona x journey stage combination. |
Content Scoring
| Tool | What it does |
|---|---|
brand_audit_content | Score content against brand rules (0-100) across multiple dimensions. |
brand_check_compliance | Quick pass/fail compliance gate before publishing. |
brand_audit_drift | Detect systematic brand drift across multiple pieces of content. |
Runtime + Utilities
| Tool | What it does |
|---|---|
brand_runtime | Read the compiled brand runtime contract (single-document brand context for AI agents). |
brand_context | Select a task-scoped slice of the runtime deterministically (task_type → sections, audience → persona match, compact budget). Returns matched selectors and explicit no-match — never silent fallback. |
brand_write | Load full brand context (visual + voice + strategy) for content generation. |
brand_export | Generate portable brand files for Chat, Code, team sharing, or email. |
brand_feedback | Report bugs, friction, or feature ideas to the brandsystem team. |
Brandcode Studio Connector
| Tool | What it does |
|---|---|
brand_brandcode_connect | Connect a local .brand/ directory to a hosted Brandcode Studio brand and pull the current package. |
brand_brandcode_sync | Pull updates from a previously connected hosted brand using sync-token-aware delta behavior. |
brand_brandcode_status | Inspect the current Brandcode Studio connection, sync history, and local package summary. |
brand_brandcode_live | Toggle connected read tools to refresh from the hosted runtime within a short cache TTL. |
Tool Flow
Tools auto-chain -- each tool's response tells the LLM what to run next:
Session 1: brand_start → brand_extract_web or brand_extract_visual or brand_extract_site → brand_generate_designmd → brand_compile → brand_clarify → brand_report
Session 2: brand_deepen_identity (interview x 6) → brand_compile (generates VIM)
Session 3: brand_extract_messaging → brand_compile_messaging (interview x 3) → brand_write
Session 4: brand_build_personas → brand_build_journey → brand_build_themes → brand_build_matrixbrand_status can be called at any point. brand_preflight runs after any content generation.
CLI Commands
The npm package ships a CLI entrypoint for setup, diagnostics, and the hosted-brand connector:
npx @brandsystem/mcp doctor # local checkup: Node, profile, .brand/ state, credential permissions, client configs
npx @brandsystem/mcp install --client claude-code # write MCP config (dry-run by default; add --write to apply, existing config backed up)
npx @brandsystem/mcp install --client cline --write
npx @brandsystem/mcp install --client cursor --profile full
npx @brandsystem/mcp inspect # version, profile, tool list, .brand/ artifact inventory
npx @brandsystem/mcp brandcode connect https://brandcode.studio/start/brands/pendium
npx @brandsystem/mcp brandcode sync
npx @brandsystem/mcp brandcode statusinstall never overwrites other servers' entries: it deep-merges, backs up the existing file first, and refuses invalid JSON. It pins the spec to the version you ran (@brandsystem/mcp@<version>), including for Codex and Cline; pass --unpinned to write the floating @brandsystem/mcp instead. For protected hosted brands, add --share-token=TOKEN.
Trusted Studio origins. Credentials (session, admin, and share tokens) are only ever sent to a trusted Brandcode Studio origin — https://www.brandcode.studio and https://brandcode.studio by default — and only an approval declared by a package from a trusted origin can mark a brand production_approved. Other hosts can still be read for public packages (https only, SSRF-guarded, no credentials), but their approval claims are ignored. For local development or staging, add origins with BRANDCODE_STUDIO_TRUSTED_ORIGINS (comma-separated; https required, except http://localhost, http://127.0.0.1, or http://[::1] when listed explicitly).
The .brand/ Directory
After running the full pipeline, your .brand/ directory looks like this:
.brand/
brand.config.yaml # Client name, industry, source URLs, session state
core-identity.yaml # Colors, typography, logos with confidence scores
extraction-evidence.json # Multi-page rendered evidence bundle (optional)
design-synthesis.json # Structured design synthesis (radius, shadow, layout, personality)
DESIGN.md # Portable agent-facing design brief
tokens.json # DTCG design tokens (compiled output)
brand-runtime.json # Compiled runtime contract (single-doc brand context)
interaction-policy.json # Enforceable rules (anti-patterns, voice, claims)
needs-clarification.yaml # Items requiring human review
brand-report.html # Portable HTML brand report
visual-identity.yaml # Session 2: composition, patterns, anti-patterns
visual-identity-manifest.md # Session 2: compiled VIM document
system-integration.md # Session 2: CLAUDE.md / .cursorrules setup guide
messaging.yaml # Session 3: perspective, voice, brand story
messaging-audit.md # Session 3: voice fingerprint analysis
brand-story.md # Session 3: compiled brand narrative
assets/
logo/
logo-wordmark.svg # Extracted logo files
illustrations/ # Brand illustrations with MANIFEST.yaml
stickers/ # Brand stickers with MANIFEST.yaml
patterns/ # Brand patterns with MANIFEST.yamlFile Details
| File | Format | Purpose |
|---|---|---|
brand.config.yaml | YAML | Project metadata: client name, industry, website URL, Figma file key, session number, schema version |
core-identity.yaml | YAML | All extracted brand data: colors (with roles and confidence), typography (with families and weights), logo specs (with inline SVG and data URIs), spacing |
extraction-evidence.json | JSON | Multi-page rendered evidence captured from representative pages and viewports. Contains screenshots, computed elements, and CSS custom properties used to ground synthesis |
design-synthesis.json | JSON | Structured design interpretation of the brand. Includes radius, shadow, spacing, layout, component, motion, and personality signals derived from evidence and current identity |
DESIGN.md | Markdown | Portable agent-facing design brief synthesized from the evidence bundle and current brand state |
tokens.json | JSON | DTCG design tokens. Includes colors and typography plus synthesis-driven radius, shadow, layout, spacing, and motion groups when available |
brand-runtime.json | JSON | Single-document brand contract for AI agents. Merges all 4 session YAMLs into flat, fast-access format. Only medium+ confidence values. Compiled by brand_compile, read by brand_runtime |
interaction-policy.json | JSON | Enforceable rules engine. Visual anti-patterns, voice constraints (never-say, AI-ism patterns), and content claims policies. Used by preflight and scoring tools |
needs-clarification.yaml | YAML | Prioritized list of items the system could not resolve confidently: missing primary color, low-confidence values, unassigned roles |
brand-report.html | HTML | Self-contained brand report. Works offline, embeds all assets inline. Paste into any AI tool as brand guidelines |
assets/logo/ | SVG/PNG | Extracted logo files. SVGs include inline path data in core-identity.yaml for portability |
Platform Setup
Codex
Use the package installer, which delegates to Codex's official MCP configuration command:
npx @brandsystem/mcp install --client codex --writeStart a new Codex task after installation. Codex CLI, the Codex app, and the IDE extension share the same MCP configuration.
Cline
Use the package installer, which delegates to Cline's official non-interactive MCP command:
npx @brandsystem/mcp install --client cline --writeStart a new Cline task, then ask: “How do I use my brand guidelines with AI?” Cline CLI and the IDE extension share the same global configuration at ~/.cline/data/settings/cline_mcp_settings.json.
Claude Code
Create .mcp.json in your project root:
{
"mcpServers": {
"brandsystem": {
"command": "npx",
"args": ["-y", "@brandsystem/mcp"]
}
}
}Cursor
Create .cursor/mcp.json in your project root:
{
"mcpServers": {
"brandsystem": {
"command": "npx",
"args": ["-y", "@brandsystem/mcp"]
}
}
}Windsurf
Create ~/.codeium/windsurf/mcp_config.json:
{
"mcpServers": {
"brandsystem": {
"command": "npx",
"args": ["-y", "@brandsystem/mcp"]
}
}
}Claude Desktop
Open Settings > Developer > Edit Config (claude_desktop_config.json):
{
"mcpServers": {
"brandsystem": {
"command": "npx",
"args": ["-y", "@brandsystem/mcp"]
}
}
}Compatibility at a glance
Three distinct ways to get your brand into an AI tool — don't conflate them:
| Path | Clients | What it takes |
|---|---|---|
| Direct local MCP (stdio) | Claude Code, Claude Desktop, Cursor, Windsurf | The .mcp.json config above — the server runs on your machine |
| Remote hosted MCP | ChatGPT (developer mode) and other remote-MCP clients | An approved Brandcode Studio brand + bearer key (pre-release, approved clients only). ChatGPT connects to remote MCP servers, not local stdio processes; OpenAI documents a secure-tunnel option for private servers |
| Runtime artifact copy | Any AI tool | Upload brand-report.html or brand-runtime.json to the conversation — no MCP connection needed |
Claude Chat (no MCP)
If you are using Claude Chat without MCP support:
- Run the pipeline in a code environment first to generate
brand-report.html - Upload the HTML file to your Claude Chat conversation
- Say: "Use this as my brand guidelines for everything we create"
The report HTML is self-contained and works as a standalone brand reference in any AI tool.
Troubleshooting
"No .brand/ directory found"
Every tool except brand_start, brand_init, and brand_feedback requires a .brand/ directory. Run brand_start first.
If you are using the hosted-brand flow instead of local extraction, brand_brandcode_connect also scaffolds .brand/ automatically on first connect.
Empty extraction (no colors or fonts found)
This usually means the website loads CSS dynamically via JavaScript. brand_extract_web only parses static CSS from <style> blocks and linked stylesheets. Solutions:
- Run `brand_extract_visual` to analyze a single rendered page with headless Chrome and computed styles
- Run `brand_extract_site` to sample representative pages across desktop and mobile and save
extraction-evidence.json - Run `brand_generate_designmd` after extraction or manual edits to regenerate
design-synthesis.jsonandDESIGN.md - Try a different page that uses more inline/linked CSS (e.g., the homepage, a blog post)
- Use Figma extraction (
brand_extract_figma) for higher accuracy - Set values manually using
brand_clarifyafter extraction
brand_start in mode='auto' already tries this visual fallback when extraction quality is low and Chrome/Chromium is available, then generates design-synthesis.json and DESIGN.md from the best available evidence.
Figma extraction fails
brand_extract_figma doesn't connect to Figma directly. It works in two phases:
- Plan returns instructions for what data to fetch (variables, styles, logo)
- Ingest processes data you pass back from the Figma MCP tools
Make sure you have a separate Figma MCP server connected (e.g., @anthropics/figma-mcp) and pass the fetched data to brand_extract_figma in ingest mode.
Logo not detected
Web extraction looks for <img>, <svg>, and <link rel="icon"> elements. If your logo is rendered via JavaScript or embedded as a CSS background, use brand_set_logo to add it manually with SVG markup, a URL, or a data URI.
"Response size exceeds 5K target" (console warning)
This is a soft warning, not an error. Some tools (brand_write, brand_deepen_identity) return rich conversation guides that exceed 5K characters. The hard limit is 50K, which triggers truncation.
Server won't start
# Verify Node.js >= 22.13.0
node --version
# Test the server manually
npx @brandsystem/mcp
# Check for port conflicts (stdio transport shouldn't have any)
# The server uses stdio, not HTTP -- it reads from stdin and writes to stdoutReporting feedback
Use brand_feedback to report bugs, friction, or ideas:
brand_feedback with category="bug", summary="Logo SVG has empty gradient stops",
detail="The extractor found the SVG structure but <linearGradient> stops have no
stop-color attributes. Logo renders as a black rectangle.",
tool_name="brand_extract_web", severity="degrades_experience"For agent telemetry, use category="agent_signal" with signal, tool_used, and signal_context. Brand context is auto-populated from .brand/config.
How It Works
Confidence Scoring
Every extracted value carries a confidence level:
| Level | Meaning | Token Behavior |
|---|---|---|
confirmed | Human-verified | Included in tokens |
high | Strong signal (e.g., Figma variable, CSS custom property named --brand-primary) | Included in tokens |
medium | Reasonable inference (e.g., most-frequent chromatic color in CSS) | Included in tokens |
low | Weak signal (e.g., color appears once in a generic property) | Excluded from tokens, added to needs-clarification.yaml |
Source Precedence
When the same brand element is found in multiple sources, the higher-precedence source wins:
figma > manual > webA Figma-sourced primary color will replace a web-extracted one. A manually confirmed value overrides both automated sources. Within the same source, higher confidence wins.
Web Extraction
brand_extract_web fetches the target URL and:
- Parses all
<style>blocks and up to 5 linked stylesheets - Extracts color values from CSS properties and custom properties
- Infers color roles from property names (e.g.,
--primary,--brand-accent) - Promotes the most-frequent chromatic color to "primary" if no explicit primary is found
- Extracts font families and ranks by frequency
- Finds logo candidates from
<img>,<svg>, and<link rel="icon">elements - Downloads and embeds logos as inline SVG or base64 data URIs
Visual Extraction
brand_extract_visual launches headless Chrome against the target URL and:
- Captures a 2x DPR screenshot of the rendered page
- Extracts computed styles from semantic elements such as body, header, hero, links, cards, and buttons
- Reads CSS custom properties from
:root - Infers likely color roles from visual context (for example, button background → primary)
- Returns the screenshot as an MCP image block so the calling agent can do qualitative visual analysis
This is the fallback path for JS-rendered apps and page builders where static CSS parsing misses key brand signals.
Deep Site Extraction
brand_extract_site extends the rendered-path beyond the homepage:
- Discovers representative pages on the same domain
- Captures desktop and mobile screenshots for each selected page
- Samples multiple instances of buttons, cards, links, inputs, sections, and other components
- Persists the results to
.brand/extraction-evidence.json - Feeds that evidence into
brand_generate_designmd/brand_compileto produce.brand/design-synthesis.jsonand.brand/DESIGN.md - Merges additional colors and fonts back into
core-identity.yamlwhenmerge=true
Use this when the homepage is not enough to understand the brand system, or when you want richer evidence before token compilation.
Figma Extraction
brand_extract_figma works in two steps to bridge between the Figma MCP and brandsystem:
- Plan mode -- Returns specific instructions for what data to fetch from Figma (variables, text styles, logo components)
- Ingest mode -- Processes the collected Figma data, maps variable names to roles, and merges into
core-identity.yamlathighconfidence
DTCG Token Compilation
brand_compile transforms core-identity.yaml into Design Tokens Community Group format:
- Colors become
$type: "color"tokens keyed by role - Typography becomes grouped tokens with
fontFamily,dimension(size), andfontWeightentries - Spacing becomes
dimensiontokens with scale values - Each token includes
$extensions["com.brandsystem"]with source and confidence metadata - Only values with
mediumor higher confidence are included
The Bigger Picture
Four verbs stack — Build, Use, Evolve, Deploy. @brandsystem/mcp owns Build. The .brand runtime is what moves between them.
Build Use Deploy
───── ─── ──────
@brandsystem/mcp ────► @brandcode/mcp ────► Every surface
(this package) (mcp.brandcode.studio) Claude Design
Claude Code
authors the ─────────► .brand/ runtime ─────────► Cursor
portable runtime (portable artifact) ChatGPT
NotebookLM
▲ Gemini
│ ...any MCP client
│
Evolve ─── Brandcode Studio
(governance promotion,
taste compilation,
memory wall)Build — this package. Extract identity. Compile governance. Produce a .brand/ directory.
Use — @brandcode/mcp hosted at mcp.brandcode.studio/{slug}. Authorized MCP clients fetch the live Full Brand Runtime by default.
Evolve — Brandcode Studio. Taste notes graduate from memory to formal governance. Anti-patterns accumulate. The runtime sharpens with every production cycle.
Deploy — the governed .brand repo travels with you. Point Claude Design at it. Compile packs for Claude Code, Chat, Gemini, NotebookLM. Every surface consumes the same runtime.
Progressive Depth
Each stage builds on the previous. Stop anywhere — you get value immediately.
| Stage | What You Get | How |
|---|---|---|
| 1. Free scan | Brand tokens + DESIGN.md + HTML report with platform setup guides | brand_start (auto) or brand_extract_web / brand_extract_visual / brand_extract_site → brand_generate_designmd → brand_compile → brand_report |
| 2. MCP depth | Figma extraction, clarification, full audit | Session 1 with brand_extract_figma + brand_clarify |
| 3. Visual identity | Composition rules, patterns, anti-patterns, VIM | Session 2: brand_deepen_identity → brand_compile |
| 4. Core messaging | Voice profile, perspective, brand story | Session 3: brand_extract_messaging → brand_compile_messaging |
| 5. Studio sync | Hosted package pull, sync history, shared distribution | brand_brandcode_connect → brand_brandcode_sync → brand_brandcode_live for Live Mode reads |
| 6. Live Use MCP | Authorized agents hit mcp.brandcode.studio/{slug} for current runtime, knowledge search, draft checks, asset fetch | @brandcode/mcp connects once with bearer-key auth; reads stay fresh across agent sessions |
| 7. Deploy to Claude Design | Claude Design grounds on the .brand/ directory natively — governance, narratives, proof points, taste notes all load without translation | Point Claude Design at a repo containing .brand/; output is on-brand from the first generation |
Stages 1–4 are the standalone local MCP flow. Open source, fully portable, no account required.
Stages 5–7 are the Deploy path — where the .brand runtime becomes shared, served live, and consumed by every generation surface. Available through Brandcode Studio and Column Five Media.
What's Portable
| Artifact | Portable? | Owned By |
|---|---|---|
@brandsystem/mcp (authoring tool) | Fully — open source, any brand | MIT license |
.brand/ directory (the runtime) | Fully — the portable artifact that travels with your brand | Client |
| Brandcode framework (schema + stances + U-mech) | Fully — universal layer imported by every brand instance | Open |
| Client claims, narratives, rules (I-content) | Per-instance — unique to each brand | Client |
@brandcode/mcp (hosted Use MCP) | Serves the runtime — authorized MCP clients connect | Brandcode |
Development
# Install dependencies
npm install
# Build TypeScript
npm run build
# Watch mode for development
npm run dev
# Run tests
npm test
# Watch mode for tests
npm run test:watch
# Type check without emitting
npm run lint
# Start the server (stdio transport)
npm startProject Structure
src/
index.ts # Entry point -- stdio transport
cli.ts # CLI entry point for brandcode connect/sync/status
server.ts # MCP server creation and tool registration (all tools, priority order)
tools/ # One file per tool
brand-start.ts # Entry point (Session 1)
brand-status.ts # Progress dashboard
brand-extract-web.ts # Website extraction
brand-extract-figma.ts # Figma extraction (plan/ingest)
brand-set-logo.ts # Manual logo add/replace
brand-compile.ts # Token + VIM + runtime compilation
brand-clarify.ts # Interactive clarification
brand-audit.ts # Schema validation
brand-report.ts # HTML report generation
brand-init.ts # Low-level directory scaffolding
brand-deepen-identity.ts # Session 2: visual identity interview
brand-ingest-assets.ts # Session 2: asset cataloging
brand-preflight.ts # Session 2: HTML compliance checking
brand-extract-messaging.ts # Session 3: voice/messaging audit
brand-compile-messaging.ts # Session 3: perspective + voice interview
brand-build-personas.ts # Session 4: buyer personas
brand-build-journey.ts # Session 4: buyer journey stages
brand-build-themes.ts # Session 4: editorial themes
brand-build-matrix.ts # Session 4: messaging matrix
brand-audit-content.ts # Content scoring (0-100)
brand-check-compliance.ts # Binary pass/fail compliance gate
brand-audit-drift.ts # Batch drift detection
brand-runtime.ts # Read compiled brand runtime contract
brand-brandcode-connect.ts # Hosted brand connect
brand-brandcode-sync.ts # Hosted brand sync
brand-brandcode-status.ts # Hosted brand status
brand-write.ts # Content generation context loader
brand-export.ts # Portable brand file export
brand-feedback.ts # Bug reports + feedback
connectors/
brandcode/ # Hosted brand client, persistence, and URL resolution
lib/ # Shared utilities
brand-dir.ts # .brand/ directory I/O (YAML, JSON, markdown, assets)
confidence.ts # Confidence scoring and source precedence
css-parser.ts # CSS color and font extraction
dtcg-compiler.ts # DTCG token compilation
color-namer.ts # Human-readable color name generation
content-scorer.ts # Brand compliance scoring engine
logo-extractor.ts # Logo candidate detection
svg-resolver.ts # SVG inlining and base64 encoding
report-html.ts # HTML report generation
vim-generator.ts # Visual Identity Manifest + system integration markdown
runtime-compiler.ts # Compile brand-runtime.json from 4 source YAMLs
interaction-policy-compiler.ts # Compile interaction-policy.json (enforceable rules)
response.ts # Structured MCP response builder
version.ts # Package version reader
types/
index.ts # TypeScript type definitions
schemas/
index.ts # Zod schemas for validation (7 schema files)
bin/
brandsystem-mcp.mjs # CLI entry point
specs/
brand-runtime-schema.md # Runtime contract documentation
interaction-policy-schema.md # Interaction policy documentation
test/
lib/ # Library unit tests (9 files)
tools/ # Tool tests (2 files: export + smoke)
server.test.ts # Server creation smoke testAgent Evaluation Suite
The eval/ directory ships fixtures, methodology, and a runnable harness measuring what matters to agents: first-tool selection on real prompts, response token budgets, envelope conformance, compliance-check accuracy against labeled cases, and whether a second agent can use a generated runtime with no explanation. Run the deterministic tier with npm run eval; the model-dependent tier (first-tool selection) is opt-in via ANTHROPIC_API_KEY. Results are published only from actual runs with stated model versions and dates — the repo ships evidence machinery, not claims. See eval/README.md.
Security
- Report vulnerabilities privately via the repository's Security page → Advisories → "Report a vulnerability". See SECURITY.md. Please don't open public issues for suspected vulnerabilities.
- Local-first by design. Extraction and compilation run on your machine. Network activity is limited to: fetching the sources you point the tools at (websites, Figma via your own Figma MCP), optional Brandcode Studio connector calls when you connect a hosted brand, and optional
brand_feedbackreports. - What gets written locally: everything lands in
.brand/inside your working directory. Connector credentials are stored in.brand/brandcode-auth.json(gitignored).
License
MIT