Raven

Raven MCP

Odin's ravens brought back knowledge of the world — Raven brings back design intelligence.

Raven is an MCP server for coding agents. Click any element in the app you have running locally and say what should change — Raven sends the agent the selector, the computed styles, and your design tokens — then audits the result for contrast, tap targets, and typography.

Raven MCP is a personal open-source project by Andrew Cunliffe. It is not endorsed by, affiliated with, or supported by Intuit Inc. or any other company referenced in its source data. See NOTICE for full attribution of upstream sources and their licenses.

What it does

Raven gives Claude access to a comprehensive design knowledge base:

• Principles — Nielsen's 10 Heuristics, all 21 Laws of UX, Gestalt principles, WCAG accessibility, typography rules, color theory, mobile UX, D4D framework, UX writing, service design, brand, color-systems (palette-size discipline), and spacing-systems (base-unit grid + scale limits)
• Patterns — Proven UI patterns for signup flows, pricing pages, navigation, dropdown/select menus, forms, landing pages, dashboards, modals, empty/error/loading states, CTAs, social proof, mobile conversion — plus content patterns (error messages, empty-state copy, notifications, form validation) and service patterns (service blueprinting, human handoff, signup-as-service, omnichannel continuity, moments of truth)
• Content systems — Voice & tone guides: Conversational Product Voice, GOV.UK, Shopify Polaris, and Atlassian
• Research — Qualitative, quantitative, and usability methods with do/don't protocols and checklists. Metrics frameworks: HEART, AARRR/Pirate, North Star Metric, conversion funnel, RICE, OKRs.
• Service design — Service blueprinting (with HTML blueprint generation — current vs. ideal state), human-handoff patterns, signup-as-service, omnichannel continuity, moments of truth / recovery, and the GOV.UK Service Standard
• Brand & visual — Logo usage (clear space, min sizes, variants, placement, restraint), gradient usage (hierarchy, palette, contrast, trend vs signature), imagery (consistency, representation, purpose), visual hierarchy, brand-as-system, and current (2026) visual-design trends
• Business — Monetization models, retention strategies, onboarding optimization, growth mechanics, and product metrics frameworks
• Tokens — Design system tokens for Stripe, Linear, and more
• Creative studio — Local-first brand profiles, asset references, character reference profiles, provider-agnostic image/video/3D/audio generation jobs, campaign plans, and transparent creative scoring. Raven does not ship media-provider credentials; set RAVEN_CREATIVE_RUNNER to route jobs to your own renderer.

Install

Local stdio (npx / from source) is the full product: 115 tools, including Grab, the pattern library, and the file-backed Taste Engine. Hosted endpoints are smaller subsets — pick one path and stick to it.

Claude Code — one command

claude mcp add raven -- npx -y raven-mcp

Prefer one Raven entry. If both a local raven and a claude.ai / remote Raven are connected, the agent sees two overlapping toolsets — disable or rename one (e.g. raven-local vs raven-cloud) so it is obvious which product you are talking to.

Manual config (Claude Desktop or team .mcp.json)

{  "mcpServers": {    "raven": {      "command": "npx",      "args": ["-y", "raven-mcp"]    }  }}

Cursor

Same mcp.json snippet as above (~/.cursor/mcp.json or project .cursor/mcp.json) runs the full local server (Grab + Taste). Hosted options:

• Public: "url": "https://mcp.ravenmcp.ai/api/mcp" — ~45 stateless tools; no Grab, no Taste.
• Authenticated Taste: "url": "https://mcp.ravenmcp.ai/api/mcp-user" — OAuth; Taste yes, Grab still local-only.

Codex

Add under mcp_servers in config.toml:

[mcp_servers.raven]command = "npx"args = ["-y", "raven-mcp"]

Codex may prompt to approve many Raven tools on first use — that is client approval policy, not a smaller Raven.

Claude Desktop — one-click extension

Prefer not to edit JSON? Download raven.mcpb and double-click it. Claude Desktop installs Raven automatically — no Node, no terminal. Package version tracks npm.

From source

git clone https://github.com/rhinocap/raven-mcp.gitcd raven-mcp && npm install && npm run build

Tools

Decision Graph

The local Decision Graph keeps three node kinds: decisions, evidence, and sources. Five edge types connect them: supersedes, scoped_alongside, supports, contradicts, and derived_from. Decision status is candidate, active, superseded, or contested; nodes are not hard-deleted.

• decision_add — add an active decision with its scope, component, rationale, and rejected alternatives.
• decision_evidence — attach quantitative or qualitative evidence to a decision.
• decision_get — return one node, its connected neighbors, and attached evidence.
• decision_list — list active, superseded, contested, or candidate decisions. Candidates are excluded unless include_candidates:true or status:"candidate" is passed.
• decision_draft — capture a decision before its rationale is confirmed.
• decision_commit — confirm a rationale and surface similar active decisions for review.
• decision_supersede — replace a decision while keeping both nodes and their lineage.
• decision_scope — narrow two active decisions so they can coexist.
• decision_history — return a supersession lineage from oldest to newest.
• ingest_transcript — store a Source node and return the extraction prompt for the calling model.
• decision_import — read local git history and matching decision documents, then return source-bound extraction prompts.
• ingest_transcript_results — turn extracted JSON into candidate decisions linked with derived_from edges.
• gap_scan — rank uncovered components, missing or thin rationales, contested decisions, and derived staleness; digest_only:true is quiet when no action is needed.

For a cold start: call decision_import → run the returned extraction prompts with a model → pass each result to ingest_transcript_results → review the candidates → call decision_commit for each decision to keep. Candidates remain available through decision_get, but default decision_list and gap_scan ignore them until commit changes their status to active.

Figma comment archives (Markdown files under figma-comments-archive/ whose first line is # Figma comments archive: <label>, with ## Thread <n> headings) are picked up by default. Their settled threads use thread-aware extraction with path#Thread <n> provenance; imported candidates still require decision_commit and are never auto-committed.

Imported provenance is checked against its Source node before evidence is attached. Git references must be a full or unique-prefix match for a commit included by that import. Document references must match the imported path, optionally followed by a line (#L12) or heading fragment. Rejected references are returned in rejected_source_refs; the candidate remains available without an evidence node.

For transcripts: call ingest_transcript → run its extraction prompt → pass the result to ingest_transcript_results → review and commit the candidates. Resolve active conflicts with decision_supersede or decision_scope, inspect lineage with decision_history, and use gap_scan for health checks.

Evidence nodes and supports / contradicts edges capture quantitative and qualitative results linked to decisions.

review_diff severity policy

review_diff is advisory by default (verdict caps at warn). Two independent, combinable opt-ins escalate matching violations to error, producing a failing CI verdict:

• fail_on — a rule allowlist. Valid rules: important, bare-hex-color, hardcoded-font-size, hardcoded-font-family, hardcoded-spacing. Start with important; add token rules once DESIGN.md tokens are mature. important findings can include intentional uses (email-client compatibility, responsive overrides), so expect to justify or restructure those hunks; token rules only fire when DESIGN.md defines tokens (checks_skipped tells you when they didn't run).
• fail_on_governed — escalates findings a recorded decision governs (lexical scope+category association, not a verified contradiction). Opt in as a team strict-mode signal.

Escalation is diff-scoped: only newly added lines can fail — existing violations don't block until a diff touches them. The applied policy is echoed back under severity_policy. Omitting both keeps the existing advisory behavior unchanged. review_diff is local-stdio only (not on the hosted remote endpoints), so wire the policy into CI via npx raven-mcp.

Archive Figma comments

Archive your Figma comment history to durable JSON/Markdown before you lose access: FIGMA_TOKEN=<pat> node scripts/figma-comments-archive.mjs --md <fileKey> The PAT needs file_comments:read. Add --resolve-nodes for best-effort node names; it also needs file_content:read, and archival still succeeds if resolution is unavailable.

Without credentials: in Figma, first show resolved comments and clear any comment filters (hidden threads won't be in what you copy — and they're unrecoverable after cancellation). Figma has no bulk "copy all comments", so select and copy the thread text from the comments panel, then run (macOS): pbpaste | node scripts/figma-comments-archive.mjs --paste design-review (the last word is your archive label — any name without spaces; add --out somedir to choose the folder). Or run the command bare and paste into the terminal, ending with Ctrl-D. Separate threads with a blank line; within a thread, an author line followed by a timestamp line ("2 days ago", "Yesterday", "Mar 4, 2026") starts each comment. Paste mode writes <label>.txt (your paste, byte-verbatim — the durable record) and always renders the readable <label>.md archive. Skim the .md against your paste: message lines that themselves look like a timestamp, or blank lines inside one comment, can shift how the .md groups things — the .txt is always exact. An existing label is never overwritten; pass --force to replace it.

Click-to-change (grab) + DESIGN.md

Grab is local-stdio only. Hosted Cursor/Claude remote endpoints do not expose Grab — click-to-change needs a loopback bridge on your machine. Use local npx / Cursor local mcp.json when you need Grab.

Raven Grab connects a local page to your agent so you can click an element, describe the change, and send its selector, computed styles, matching DESIGN.md tokens, and token choices back to the session. The bridge runs on loopback and the returned script tag carries the capability key required by its routes. Computed styles are editable inline, and edits are sent to the agent as styleEdits.

Setup takes under a minute:

1. Start your local dev server.
2. Call start_grab_session with proxy_target set to the local server URL. path to a DESIGN.md is optional when proxy_target is set (Raven creates a minimal temp DESIGN.md); required when you only inject the script without a proxy.
3. Open the returned bridge URL. The overlay is already included on HTML pages served through it.
4. Click elements and enter the changes you want in the Grab panel.
5. Call get_grabbed_elements to receive the queued selections and instructions (draining frees queue capacity for later sends).

For a page you control, you can omit proxy_target and paste the returned <script> tag into the page instead.

Use read_design_md to inspect a DESIGN.md file and its flattened token index, init_design_md to create one from a stored Raven system, a blank template, or a getdesign.md starter, and update_design_md to set, rename, or remove one token without rewriting the rest of the file.

Pattern library — keep what you grab, then translate it

proxy_target also accepts a third-party URL, so you can grab from any site you are allowed to view, not just your own dev server. What you grab is otherwise gone when the tab closes, and it arrives as another site's literal values. Four tools close that loop:

• capture_reference — persist a grabbed selection under ~/.raven/references: selector, computed styles, hover/focus states, bounding rect, truncated HTML, your own note, and tags. One JSON record per capture, so grabbing the same element twice keeps both. It also renders the captured element back into a PNG beside the record, because nobody can pick a pattern out of a style map. That render is offline — every external request is aborted, so a stored reference never reaches back out to the site it came from — and it runs with scripting disabled, so a script in a captured element cannot execute. The record says so: fidelity: "offline".
• search_references — find it again later by free text, host, owner, or tags, with a per-result score and a why naming the fields that matched. Each result carries a display object holding the credit line and the image path together, so a consumer reaching for the picture carries the attribution out with it. Looking is not copying: a result reports html_available but omits the captured markup, because browsing a corpus of other people's work should not hand back their markup as a side effect of looking at it. Pass include_html: true when you actually mean to read the structure — the response then names whose markup it contains. Everything a browse is for is in the default result: the picture, the selector, the rect and the computed styles.
• map_reference_to_tokens — translate the captured literals onto your DESIGN.md tokens, so the code an agent writes uses your type ramp and palette instead of pasted values. Every binding carries the resolved value and CSS variable alongside the token name — a name alone is not something you can write into a stylesheet — and an aliased token resolves to the literal at the end of its $ref chain. Pure and deterministic: no model, no network. It reads the stored styles directly, so the whole show-it-then-translate-it path runs without the markup ever leaving Raven.
• forget_references — remove a single reference by ref_id, or every reference from a host (subdomains included). Takes the PNGs with it. Destructive and permanent, so the host sweep refuses to run without confirm: true and tells you how many records that would remove first.

How the mapping decides, because a wrong binding is worse than a stated gap:

• Colour matches on RGBA distance, not RGB — the same hex at a different opacity is a near miss, not an exact hit. Hex, rgb()/rgba() in both comma and space form, hsl(), and the CSS named colours all resolve; a syntax the matcher cannot read (oklch(), lab()) says so by name instead of reporting your palette as empty.
• Lengths normalize to px at a 16px root. Percentages and viewport units need a containing size and are returned as gaps with that reason, never converted on a guess.
• Family before proximity. A property that belongs to a token family binds inside it: padding-top takes a spacing token even when a type token is numerically closer, and line-height takes the leading token over an equally-exact size token. When no token in the right family is close enough, the result is a gap that names the cross-family near miss ("the closest token by value is space.4 (16px), but it belongs to a different family") rather than binding font-size to your spacing ramp.
• Ties break on distance, then family fit, then shortest and lexicographically-first token path, so the same inputs always produce the same binding regardless of token order.
• Broken $ref chains and cycles in your DESIGN.md come back in diagnostics even when every property still found a match — a defect in your own token file is reported, not swallowed.

Respect the source. Grab from sites you are permitted to access; the tools never bypass a paywall or a login, and owner: "third-party" is recorded on every capture.

Attribution and takedown. Every third-party record keeps the URL, host, app name and capture date it came from, and search_references derives a credit line from them on read — so the credit cannot go stale, and it travels with the picture rather than beside it. Raven claims no ownership of anything you capture.

Your corpus is local: it lives in ~/.raven/references on your own machine, and this project hosts no copy of it. So a takedown is something you run, not something you request — if a rights holder asks you to remove their material, forget_references with their host removes every record from that host and every subdomain, and the images with it:

forget_references({ host: "example.com", confirm: true })

It reports what it removed, what it could not read, and anything it tried to remove and failed — those are three different answers and it does not collapse them into one. A removal that fails part-way leaves the record in place rather than the picture, so running it again finds and finishes what was left. The confirmation prompt names the exact records it would take, and passing those ids back as expected_ref_ids pins the removal to them — anything captured in between is reported rather than swept up. If you believe this project itself is distributing your material, open an issue at https://github.com/rhinocap/raven-mcp/issues.

One boundary worth stating plainly: while the bridge is proxying a third-party site, that page is served from the bridge's own origin, so scripts on it are same-origin with the Raven overlay and can read your DESIGN.md token names and values. Raven withholds the DESIGN.md file path and every authoring route (layer moves, template and component writes, batch commits) for the duration of a proxy session, but proxy sites you would be comfortable showing your token list to.

Creative studio

Raven now covers the creative-production workflow around media generation without copying or depending on any closed vendor. The tools are orchestration primitives:

• Store brand kits locally with create_brand_profile.
• Register product photos, logos, references, or URLs with register_creative_asset.
• Create character/identity reference sets with create_character_profile.
• Generate provider-ready payloads with create_generation_job.
• Build full campaign shot lists with plan_creative_campaign.
• Score creative concepts with score_creative.

By default, jobs are saved as local draft payloads under ~/.raven/creative (override with RAVEN_CREATIVE_HOME). To run real media generation, set RAVEN_CREATIVE_RUNNER to an executable that reads one job JSON object from stdin and returns JSON on stdout. That runner can call any provider you choose; Raven never stores API keys in source.

iOS / SwiftUI audits

Raven audits native iOS apps against the Apple Human Interface Guidelines, not web/CSS conventions. None of the web-only rules (lang, title, flex-wrap, clamp, max-width, CSS custom properties, bare hex) run on iOS input — and get_checklist/get_principles take platform: "ios" to return HIG items (Dynamic Type, 44pt targets, SF Symbols, safe areas, dark-mode parity, App Review privacy) instead of the web set.

• audit_swiftui — paste SwiftUI source (source: a string or array of files). Statically flags hardcoded .font(.system(size:)) below ~13pt, tiny semantic fonts (.caption/.caption2), hardcoded Color(red:green:blue:)/hex literals (vs. asset-catalog or semantic system colors), interactive frames under 44×44pt, and ad-hoc spacing off the 4/8-pt grid. Rewards semantic Dynamic Type fonts, semantic system colors, SF Symbols, and flexible frames. Pass the optional accent_color_contents (the raw AccentColor.colorset/Contents.json) and it verifies the accent color actually defines components — catching an empty/undefined AccentColor that would silently fall back to system blue.
• audit_ios_screen — the iOS analog of audit_layout. Call with no args for the expected snapshot shape and how to capture it (Accessibility Inspector / XCUITest). Call with { elements: [{ label, rect, role, fontPt, fgColor, bgColor }], viewport } (plus an optional base64 screenshot) to score 44×44pt touch targets, contrast (with iOS secondaryLabel/tertiaryLabel treated as platform-standard — a warning, not a hard fail), and visual rhythm (alignment, gap consistency, optical balance).
• audit_ios_privacy — the "no sketchy issues" gate. Reads info_plist or an Expo app_json (managed RN apps have no Info.plist) plus optional privacy_md, entitlements, and source. Flags NS*UsageDescription strings that are vague or contradict the code (e.g. an NSHealthUpdateUsageDescription write claim that requestAuthorization(toShare: []) never fulfills), unused entitlements, Android permissions (Expo), ATS cleartext exceptions, secrets/keys shipped in the bundle or app.json extra, and default data-egress paths not disclosed at the point of choice (a pre-selected "Recommended" option that silently sends personal data to a hosted server).

All three return the same shape as audit_page — score, grade, summary, passes, errors, warnings, fix_priority (with audit_ios_screen adding a metrics block).

One command: node scripts/ios-audit.mjs <app-dir> [--snapshot snap.json] [--md report.md] discovers all the inputs and runs all three tools with an aggregated report.

React Native / Expo audits

Anyone building a React Native or Expo app gets the same treatment. RN renders to native iOS + Android widgets, so audit_ios_screen already scores its rendered output (an accessibility snapshot is platform-level); audit_rn covers the JSX/StyleSheet source — the RN analog of audit_swiftui — graded against the iOS HIG + Android Material conventions RN has to satisfy on both platforms. get_checklist/get_principles take platform: "react-native".

• audit_rn — paste RN source (source: a string or array). Flags touchables (Pressable/Touchable*) missing accessibilityLabel/accessibilityRole, touchables under 44pt with no hitSlop, allowFontScaling={false} (silently breaks Dynamic Type), fontSize below ~13, screens with no SafeAreaView/useSafeAreaInsets, and — for multi-mode apps — hardcoded colors with no useColorScheme/Appearance. Pass color_scheme: "dark"/"light" (your Expo userInterfaceStyle) and the dark-mode check is suppressed for intentionally single-mode apps. Rewards SafeAreaView, hitSlop, Platform-aware code, and a theme.
• audit_ios_privacy also accepts an Expo app_json — it audits expo.ios.infoPlist, Android permissions, plugins, and scans expo.extra/config for secrets and Google API keys.

One command: node scripts/rn-audit.mjs <app-dir> [--snapshot snap.json] [--md report.md] discovers screens + app.json (reading userInterfaceStyle so dark-only apps aren't false-flagged) and runs everything.

Responsive visibility audits

audit_responsive_visibility renders a page at multiple breakpoints (default: 390px mobile, 768px tablet, 1440px desktop, 2160px ultra-wide) and flags content elements that are visible on desktop but hidden on mobile — catching the "vanishes on mobile" bug class. Each flagged element is categorised as likely-oversight (content that shouldn't be hidden) or intentional (decorative elements). Detects hiding via CSS (hidden, display:none, opacity:0, visibility:hidden) and responsive Tailwind classes (hidden md:block, etc.).

Usage:

• audit_responsive_visibility(url) — render at default breakpoints and flag mismatches.
• audit_responsive_visibility(url, [390, 768, 1440]) — custom breakpoints.
• Optional viewportHeight (default: 900px) for tall content.

Returns flagged elements with selector, hiding class, visibility at each breakpoint, and category.

Contrast audits

audit_contrast computes WCAG contrast for every text element, reporting a tri-state status: pass, fail, or indeterminate. Determinate rows include ratio, aa, aaa, and delta_to_aa; indeterminate rows keep required_aa but publish those four metrics as null. Gradient and layered backgrounds expose effective_bg plus ratio_min / ratio_max when a trustworthy candidate range exists, and results summarize indeterminate_bg_rows / indeterminate_bg_count separately from AA failures.

Real-backdrop compositing applies to URL mode. Raven walks the rendered DOM ancestor chain, composites parseable colors and gradient layers in CSS paint order, samples gradient interiors, and normalizes modern computed colors through the browser canvas. This intentionally stops at the DOM-ancestor ceiling: opacity, display:contents, positioned transparent chains, photos, unsupported layers, and cross-stacking-context sibling backdrops are reported indeterminate; Raven does not pixel-sample across stacking contexts. Snapshot mode retains the pre-existing supplied-bgColor / over-white model and announces that scope in mode_note.

Usage:

• audit_contrast(url) — render a live page and audit all text.
• audit_contrast(dom_snapshot: [{ selector, color, bgColor, fontPx?, bold?, text? }]) — audit a pre-captured snapshot (useful for dynamic or cookie-protected pages).

Returns all text rows with status, determinate failures with delta-to-pass, effective background evidence/ranges, and separate indeterminate summaries. suggest_contrast_fix accepts only determinate failing rows; indeterminate or null-ratio evidence is skipped rather than converted into a fake color recommendation.

WCAG math: Contrast ratio uses linearised luminance (WCAG 2.1 § 1.4.3) — black-on-white is exactly 21, white-on-black is exactly 21. Large text (18.66pt+ bold or 24pt+) needs only 3:1 / 4.5:1 AAA; regular text needs 4.5:1 / 7:1.

Headless browser audits

audit_page can render a live URL in headless Chromium, scroll to settle reveal-on-scroll elements, and play preload=none videos before capturing — preventing false "blank section" reports caused by whileInView states that haven't fired yet.

Usage:

• Static HTML mode — pass html string for immediate static analysis (existing behavior, no change).
• Rendered URL mode — pass url (full HTTP/HTTPS URL). Raven launches Chromium, renders the page, optionally scrolls, and audits the live DOM.
  • scroll_settle: true — scroll from top to bottom in viewport-height steps with a short pause at each step so IntersectionObserver / whileInView thresholds can fire, recomputing page height as lazy content appears and requiring consecutive stable-height observations at the bottom. Raven then waits for finite animations triggered anywhere in the walked page before returning to the top and re-settling for capture. Smooth scrolling is temporarily neutralized. The whole walk/settle is capped at 4s; if a very long, continuously growing, or still-animating page cannot fully settle, Raven emits a capture warning rather than silently trusting the result, and animationsSettled remains false. Unloaded videos (preload="none") are played to detect if they render blank.
  • Entrance-animation settle (always on) — before extracting content or screenshotting, Raven polls document.getAnimations() until no finite animation in or near the viewport is still running (infinite spinners/loops are ignored), capped at 3s by default. Library callers can set animation_settle_timeout_ms (hard maximum 10s). Pages whose heroes enter via animation-delay + backwards fill are captured settled, not blank or mid-flight; animationsSettled in the capture metadata reports whether quiescence was reached.
  • False-blank detection (always on) — at capture time Raven measures text-bearing leaf nodes and media/content elements across the captured page. If more than 30% are effectively invisible through opacity:0, visibility:hidden, or a fully transparent text color, capture_warnings includes reveal-gate-false-blank: … so capture-backed audit callers know the rendered audit may be untrustworthy.
  • viewport: { w, h } — set the render viewport (default: { w: 1440, h: 900 }).

Video artifacts detection: If any <video> with preload="none" (or missing preload) renders with readyState < 2 (i.e. would show a black box in a screenshot), Raven flags it as an unloaded-video-artifact in the result. This is informational — not a pass/fail — since preload=none is often intentional. On cookie-protected hosts, video requests may fail because iOS/Android media daemons don't send cookies; Raven notes this to help you troubleshoot (e.g. disable deployment protection, use a token-based bypass).

Adversarial verification: Set adversarial_verify: true to independently re-check each finding against the live DOM using a different method. Findings are tagged:

• confirmed — the finding is real on the live page (e.g. missing <title> in the rendered DOM)
• likely-artifact — the finding is an artifact of the static audit method (e.g. a <video preload="none"> rendered blank, which is expected behavior, not a missing resource)
• inconclusive — the finding cannot be independently verified (e.g. aggregate rules like color-palette size)

The result includes adversarial_verification: { debunked_count, confirmed_count, inconclusive_count }, where debunked_count is the number of likely-artifacts. This surfaces false positives so you only fix real issues. Backwards-compatible: when adversarial_verify is absent or false, the output is identical to prior versions.

Setup: First time only, run npx playwright install chromium to download the browser binary. If the binary is missing when you call audit_page with url, you'll see a clear instruction to run the install command.

Before/after design diffs

evaluate_design can now accept base64-encoded PNG screenshots to measure whether a fix actually changed the rendered output.

Usage:

• Pass before_screenshot and after_screenshot (both base64 PNGs, with or without the data:image/png;base64, prefix).
• Raven returns fix_confirmed: true if the images differ by > 0.1% of pixels (accounting for jpeg/PNG decode variance).
• changed_ratio — exact fraction of pixels that changed (0–1).
• changed_region — bounding box { x, y, w, h } of the changed pixels (null if no changes detected).
• dimensions — image-derived measurements (canvas size, brightness, color shift) as context, with the caveat that these are pixel-level proxies, not Raven principle scores.

When before/after screenshots are provided alongside a description, evaluate_design returns both the principle-based evaluation and the pixel diff. When screenshots are provided without a description, the evaluation gracefully skips the principle search and returns the diff only. Backwards-compatible: without screenshots, the tool behaves identically to prior versions.

Close the token-polish loop against a real git worktree:

npx raven-polish --apply --verify "npm test"

The CLI is dry-run by default (or node scripts/raven-polish.mjs in this repo); it exits 1 when polish is proposed or any finding has severity error, and exits 0 only when there is nothing to polish and no errors. Pass --range main...HEAD to review committed work. It checks a proposed patch before applying it, then runs review_diff again on the real repository state. For CI, copy the example workflow into .github/workflows/.

Release updates

Raven ships new principles, patterns, and brand systems regularly. For one email per minor/major release (patches stay quiet):

• Web: ravenmcp.ai/#updates — 10 seconds, one email field.
• In-product: ask Claude "register me for Raven updates at [email protected]" — Claude calls raven_register and you're in.

No marketing, unsubscribe anytime. Powered by Resend.

After you upgrade

Claude Code snapshots the tool list when the MCP server connects. After upgrading, restart the session or use /mcp to reconnect and see new tools.

Claude Desktop snapshots the tool list when the MCP server connects. After upgrading, restart the app to see new tools.

Codex CLI also requires a per-tool approval_mode entry in ~/.codex/config.toml; without it, calls to new tools are cancelled. Run node scripts/sync-codex-approvals.mjs to see what is missing, then add --write to append the entries.

The appended entries auto-approve those tools, so review the printed list; newer Codex versions can alternatively set a server-level default_tools_approval_mode.

Start every project calibrated

Taste is per-surface: the same designer wants monochrome one-accent rules enforced on their portfolio and none of them on a product site, with a slightly different voice on each. The Taste Engine handles this with a kickoff interview (once per project — every question skippable, most with plain-language multiple-choice options, from navigation pattern to aesthetic family to specialty libraries) whose answers persist as a surface binding that every future audit applies automatically. And when generated work misses, mode:'refine' turns that dissatisfaction into a re-interview against the stored binding instead of a dead end.

Starting a brand from nothing — no product, no palette, just an idea? The same interview is step one of a full genesis flow (invent the brand in conversation, generate assets with your own image tools, approve a mood board, land a design system in DESIGN.md): docs/brand-genesis-flow.md.

Raven ships this flow in its MCP server instructions, so agents that honor server instructions (Claude Code, Claude Desktop) run the interview at project kickoff on their own: get_taste_interview → ask the user → bind_taste_surface → done. If your client doesn't surface server instructions — or you want the ritual to be non-negotiable — add one line to the project's CLAUDE.md / AGENTS.md:

Before the first design/UI/copy work in this repo, run Raven's get_taste_interview(profile <name>, project <repo-name>); if existing_binding is null, ask me itsquestions and persist with bind_taste_surface. Pass project:'<repo-name>' on everyaudit_taste after that.

Already-calibrated projects cost one cheap call (existing_binding comes back non-null and the agent proceeds). Uncalibrated audits still work — scoped rules just demote to warn and the result carries a calibration_hint — so calibration is never a wall, only a sharpener.

Learning loop

Raven keeps a small local-only log of how you use it so you (and Claude) can spot which patterns you build most often and which gaps show up again and again.

• Location: ~/.raven/usage.jsonl (override with RAVEN_USAGE_LOG=/path).
• What's written: tool name, timestamp, elapsed ms, and a tiny insight object — audit score/warning rule names, pattern type, brand company name, search layer. Never the HTML you audit, never prompt text, never brand copy.
• What's never written: raw page bodies, client content, your work product.
• Disable entirely: RAVEN_NO_USAGE_LOG=1.
• Reflect: ask Claude "what have I been using Raven for?" and it will call raven_reflect, which reads the log locally and summarizes the last N days — most-used tools, recurring audit warnings (likely knowledge gaps), patterns you request most, design systems you reach for.

Nothing is sent to a remote server. If a recurring gap is worth turning into a new Raven principle or pattern, you file an issue by hand — the automated pipeline at github.com/rhinocap/raven-mcp handles it from there.

Your data on the hosted server

Raven also runs a hosted remote MCP server with two endpoints. The anonymous endpoint is stateless — no store is ever attached to it, so nothing you send is written anywhere. The authenticated endpoint (used for the Taste Engine's cross-session profiles, surface bindings, and decisions) keys everything to your account and stores it in Upstash Redis, namespaced under your verified user id; your bearer token itself is never persisted.

You can erase all of it at any time with the delete_taste_data tool (confirm: "DELETE") — it removes every key under your namespace and confirms nothing remains. Full details, including the exact key layout and the rate-limit counters that aren't part of your data, are in docs/remote-mcp-privacy.md.

Development

npm run dev    # Run with tsx (hot reload)npm run build  # Compile TypeScriptnpm start      # Run compiled output

License & attribution

Raven MCP is released under the Apache License, Version 2.0 — Copyright (c) 2026 Andrew Cunliffe.

If you fork, embed, or redistribute Raven (in whole or in part), retain the Apache-2.0 license notice, the LICENSE file, and the NOTICE file. If you ship Raven inside another product, include attribution to "Raven MCP — https://ravenmcp.ai" in your acknowledgements.

Raven's knowledge base references work from many third-party sources — Nielsen Norman Group, primary academic literature for the UX/psychology laws (Fitts, Hick, Miller, and others), Gestalt principles, WCAG (W3C), plainlanguage.gov (public domain), GOV.UK (Open Government Licence v3.0), Shopify Polaris, Atlassian Design, and others. Each entry carries a sources URL field. See NOTICE for the full list of upstream sources and license terms; some carry their own conditions beyond Apache-2.0.

This is a personal project. It is not endorsed by Intuit Inc. or any other company referenced in its source data.

Data structure

All knowledge lives in src/data/ as static JSON files:

src/data/  principles/      # Nielsen, Laws of UX, Gestalt, accessibility, typography, color, mobile, D4D  patterns/        # signup, pricing, nav, forms, landing, dashboard, modals, empty/error/loading, CTA, social proof, mobile  business/        # monetization, retention, onboarding, growth, metrics  tokens/          # registry.json + systems/ (stripe, linear, vercel, …)  content/         # voice & tone: Conversational Product Voice, GOV.UK, Shopify Polaris, Atlassian    systems/       # registry.json + brand-voice JSONs (conversational-product-voice, gov-uk, polaris, atlassian)    principles/    # UX-writing principles (clarity, active voice, error anatomy, …)    patterns/      # copy recipes for errors, empty states, notifications, form validation  research/        # study protocols + metrics frameworks    principles/    # research fundamentals (method match, bias, sample size, ethics, triangulation, …)    methods/       # qualitative, quantitative, usability    frameworks/    # HEART, AARRR, North Star, conversion funnel, RICE, OKRs  service-design/  # service-level principles + patterns + frameworks    principles/    # Stickdorn, Shostack, peak-end, moments of truth, handoff    patterns/      # service blueprinting, human handoff, signup-as-service, omnichannel, moments of truth    frameworks/    # GOV.UK Service Standard (14 points)  brand/           # brand & visual design    principles/    # logo, gradient, imagery, hierarchy, brand-as-system    trends/        # 2026-current.json