Inspo

io.github.NutlopePublished by GitHub account Nutlopev0.1.16Updated Nov 2, 2026

Real production websites for coding agents: palettes, fonts, fold autopsies and reference JSX.

VerifiedStreamable HTTPWeb executableDeveloper ToolsMedia & Design

Overview

AI-generated overview

Lets a coding assistant study real production websites — palettes, fonts, layouts and reference JSX — before it writes UI.

What you can do
  • Search and read captured website screens, with palettes, fonts, components and image URLs
  • Retrieve a screen's full record, its DESIGN.md design system, and its captured page flow
  • Find similar screens, sites near a hex colour, examples of a named macrostructure, and compare 2-4 sites
  • Browse editor-curated collections and list the accepted filter values
  • Find real sites using a given UI component and list canonical reference components
  • Fetch copy-pasteable JSX for a reference component, or run one recommend call on a brief for a full design packet
When to use it
Useful when an assistant is building or restyling a web UI and needs grounded design references: picking a layout archetype, matching a brand colour, or copying a proven hero, pricing or footer pattern instead of inventing one.
Requirements
Either the hosted Streamable HTTP endpoint or the npm package inspo-mcp run locally with Node.js. Optional environment variables: INSPO_CATALOGUE_URL, INSPO_IMAGES, INSPO_MAX_TOKENS, INSPO_PROFILE. Optional TOGETHER_API_KEY enables embedding-based semantic search; other tools work without it. No account or authentication is declared.
Before you install
Tools are described as read-only, but get_design_system with live=true fetches the source site over the network. The hosted endpoint logs tool name, success and duration per call. Reference JSX and screenshots come from third-party sites and remain their designers' work, so check licensing before shipping copied code.

Installation

In SourceWeft

  1. Open Inspo in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Web executable via Streamable HTTP. Remote servers run from the web runtime once configured in a workspace.

Use with an agent

Paste this prompt into Claude Code, Codex, Cursor or another agent — or into SourceWeft chat. The agent reads this server's install guide, shows you its publisher, endpoint and sign-in, and adds it to your client once you agree. Keys never go into a command.

Read https://sourceweft.com/mcp/io.github.Nutlope%2Finspo/install.md and connect the MCP server it describes to my client. Before changing any config, show me its publisher, endpoint and sign-in method and wait for my OK. Never put an API key in a command or in a file I would commit; ask me where it should come from.

Read the install guide the agent follows

Other MCP clients

Add this to your client's mcpServers config.

{
  "mcpServers": {
    "inspo": {
      "type": "http",
      "url": "https://inspomcp.dev/api/mcp"
    }
  }
}

README

[Inspo]

Inspo

Real production websites your coding agent can study, over MCP, before it writes UI. inspomcp.dev

With and without

Same brief, same model, no design skill on either run. The only difference: the page on the right was built with Inspo MCP on the session.

[The bare agent's page: dark theme, neon accent, gradient app mock] [The page built with Inspo: warm paper, editorial type, detailed product mock]

That gap is the product. Agents have tools but not taste, and the open web already holds every reference they need: Inspo assembles, tags, and addresses it.

What's inside

  • 2,320 pages across 832 real sites, each captured on desktop and mobile, with traced palettes, real fonts, detected tech, and a fold-by-fold autopsy on every page.
  • 68 canonical reference components (heroes, pricing, footers, navs, and more), each demonstrating one named macrostructure, with copy-pasteable JSX.
  • A DESIGN.md for every site, extracted from its DOM: semantic palette roles, type ramp, spacing scale, CSS variables, container width.

Fifteen tools expose all of it; recommend(brief) composes most of them into one call. The full tool table lives in apps/mcp/README.md.

Tech stack

  • TypeScript MCP server with two transports: stdio via npm, and Streamable HTTP served by the site's own Vercel deployment. No database at query time; the catalogue ships as a static seed.
  • Next.js 16 + Tailwind v4 gallery: the archive browser and the curator dashboard.
  • Playwright capture worker: desktop and mobile shots, palette, type ramp and CSS variable extraction, tag and autopsy passes via Together AI.
  • Together AI embeddings behind semantic search and find_similar.
  • A smoke harness that boots the server in-process and exercises the core tools.

Cloning & running

  1. One command, any client: npx -y inspo-mcp install (detects Claude Code, Cursor, Codex, VS Code, Windsurf, Zed, Claude Desktop and writes the config; --dry-run shows the plan)
  2. Or point a client at the hosted endpoint yourself: claude mcp add --transport http inspo https://inspomcp.dev/api/mcp
  3. Or clone: git clone https://github.com/Nutlope/inspo.git && cd inspo && pnpm install
  4. Gallery: pnpm dev, then open localhost:3000
  5. Dev loop: pnpm --filter @inspo/mcp test

Per-client snippets (Cursor, Windsurf, Zed, Claude Desktop) are on the MCP page; the self-host runbook is DEPLOY.md.

Roadmap

  • show the desktop and mobile capture side by side on every screen page
  • backfill component crops so find_components returns a crop for every hit
  • paginate the list tools instead of capping them
  • re-shoot the mobile set at retina scale
  • Accounts and paywalls are skipped on purpose: the archive is more useful free.

Want a site in the archive? Append it to apps/worker/src/seed-urls.ts and open a PR. The bar: does it make the archive better for someone building a website?

Security

Every tool is read-only. The one path that touches the outside world - get_design_system(live:true), which supplements thin captured tokens from the screen's own source - passes each URL and redirect through an SSRF guard: public named hosts only, ports 80 and 443, byte-capped body, no JS execution. The hosted endpoint logs tool name, success and duration per call, never IPs or query text.

MIT, copyright Together AI and contributors. The screenshots remain the work of their designers: every screen credits and links its source, and takedowns are honoured at /dmca.

Source: README.md at commit 5d2f4ea

Tools

15

search_screens

Search the archive of real website screenshots. Returns screens with palette, fonts, components and image URLs.

ParametersmodevibecolorlimitquerystyledetaildeviceindustrypageTypeaccentHuemaxTokens+3 more

get_screen

Fetch the full record for one screen by slug - every viewport variant, palette, fonts, tech stack, designer credit.

Parametersslugrequired

get_design_system

The full DESIGN.md for one screen - real fonts, frequency-ranked palette, CSS variables, detected tech, colour roles, type ramp. Thin captured tokens are supplemented by a live fetch of the source; `live=false` skips that.

Parametersliveslugrequired

find_similar

A screen's nearest design neighbours - ranked by design-similarity embeddings where available, else by macrostructure / industry / style overlap. One result per site.

ParametersslugrequiredlimitdetailsameSitemaxTokens

compare

Compare 2-4 captured sites across palette, typefaces, macrostructure, mode, style tags, type/spacing/radius scales and container width. Returns per-site breakdowns plus what they share - triangulates a house style from a few references.

Parametersslugsrequired

find_by_color

Real sites whose extracted palette sits near a hex colour (OKLAB distance, so 'close' means the same colour family, not just the same hue). Use when a brief names a brand colour.

ParametershexrequiredlimitdetailmaxTokenstolerance

find_examples_for_macrostructure

Real sites exemplifying one of the 19 named macrostructures. Call it at the macrostructure-pick step. Accepts slugs ('bento-grid') or display names ('Bento Grid').

ParametersnamerequiredlimitdetaildevicemaxTokens

list_collections

List every editor-curated issue. Use this for thematic browsing - e.g. all 'Editorial Layouts' or 'Dark Product Pages'.

Takes no parameters.

get_filters

Every value the filters accept: styles, industries, macrostructures, modes, vibes, colors, page types, component types. Zero input. Call it when unsure - an out-of-vocabulary value errors.

Takes no parameters.

get_site_pages

A site's captured pages as an ordered flow (landing → pricing → features → ...) with titles, northstars and thumbnails - how a real product sequences its pages. No arguments returns the directory of flow-capable sites. Accepts a site or screen slug.

ParameterslimitsiteSlugmaxTokens

find_components

Real sites featuring a specific UI component (hero, pricing, features, cta, nav, footer, testimonial, logo-cloud, faq, stat), with a per-element crop where one exists. For copy-pasteable code use get_reference_jsx.

ParametersmodetyperequiredvibecolorlimitstyledeviceindustrypageTypemaxTokensmacrostructure

get_collection

Fetch one issue by slug, including the editor's blurb and ordered screen list.

ParametersslugrequireddetailmaxTokens

find_reference_components

The canonical reference components as an index: id, label, macrostructure and a note for each hero / pricing / cta / nav / footer archetype. Pick one, then get_reference_jsx(type, id) for its source.

Parameterstypemacro

recommend

Orchestrator - start here for a brief. One call returns a macrostructure pick, 5 real exemplars, matching reference components (source via get_reference_jsx), a palette suggestion, and an `evidence` packet measuring the genre's paper band / display class / accent hue. Pass a macrostructure to skip the pick step. No LLM call.

ParametersmodevibebriefrequiredcolordetaildevicepageTypemaxTokensmacrostructure

get_reference_jsx

The full canonical JSX source for one reference component, including the header stamp naming its macrostructure / theme / states. Copy-pasteable; tweak tokens to the target brand.

Parametersidrequiredtyperequired

Version history

1
  1. v0.1.16LatestSep 16, 2026

Score

100% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 daysRequired25 pts
  • Auth declaredRequired10 pts
  • Tools listedRequired15 pts
  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts