Arche MCP documentation — Arche Directory

What it is

Arche Directory is a hand-picked reference library of carefully made websites. Arche MCP is a remote MCP server that gives an AI client the landing pages and design studio sites in it as references: screenshots, typefaces, colours and design notes, and for each measured site its design system, the animations its live page runs (as CSS, GSAP or Motion code) and its sections. It also plans a page from a brief and runs a craft check of a page you built, set beside a reference. It is for designers and developers who build pages with an agent and want them to start from real, well-made sites and end up with something of their own.

Every tool is read-only: the server never changes anything in the archive or in your account. Personal portfolios are not shared with agents unless their owner asked for it.

Connect

Address: https://arche.directory/mcp
Transport: Streamable HTTP.
Sign-in: OAuth 2.1 with dynamic client registration; you sign in with Google. Metadata at /.well-known/oauth-protected-resource/mcp.
Claude (web, desktop, mobile): Settings → Connectors → Add custom connector, paste the address, Add.
Claude Code: claude mcp add --transport http arche https://arche.directory/mcp, then /mcp to sign in.
Cursor: Settings → MCP → Add new MCP server, or the Add to Cursor button on the MCP page.
ChatGPT: Settings → Apps & Connectors → developer mode → Create, paste the address, authentication OAuth.
Codex: codex mcp add arche --url https://arche.directory/mcp && codex mcp login arche

Tools

A site can be named by its slug, its name or its domain in every tool that takes slug.

search_sites Search the archive: Finds sites by free text and filters; each result has its screenshot URL, typefaces, colour, stack and a one-line summary.; Inputs: query, category (landing / studio), tags, font, theme (light / dark), with_video, motion, limit (up to 40)
get_site Get one site: Everything Arche knows about one site: summary, design notes, typefaces, colour, stack, techniques, screenshots, motion preview, collections and similar sites.; Inputs: slug: slug, name or domain
get_brief Design reference brief: A ready-to-paste design reference for one site, and how to use it as a starting point rather than a template.; Inputs: slug
get_design_system Design system: The design system measured on the live page: colour tokens with roles and contrast, type scale, spacing, radii and the primary button.; Inputs: slug, format (markdown / css / tailwind)
get_motion Motion recipes: The animations the live page runs, with their measured durations, easing and stagger, written as code.; Inputs: slug, library (css / gsap / motion / all)
search_sections Search sections: Finds sections of real pages (heroes, logo rows, features, pricing, testimonials, FAQs, CTAs, footers…), each as an image with its measured facts.; Inputs: query, type, theme, motion, category, limit (up to 24)
get_section Get one section: One section of a site: its image, measured facts, the design tokens that apply and the code for its motion.; Inputs: slug, section (its number)
plan_page Plan a page: A page structure for a brief, drawn from sites that fit it: which sections, in what order, how tall and dense, real sections to look at and one design system to start from.; Inputs: brief, category, theme, like (sites whose feel you want)
check_page Craft check beside a reference: A craft check of a built page, set beside a site: type scale, colour and contrast, spacing, section rhythm, first screen and motion, with both numbers wherever they differ. Differences are not errors; the reference is not a target.; Inputs: slug, measured (the JSON from check.js, see below)
similar_sites Similar sites: Sites that share a category and tags with the given one.; Inputs: slug, limit (up to 12)
list_tags Tags and collections: The vocabulary for filtering: categories, every tag by facet with counts, and the curated collections.; Inputs: none
saved_sites Your saved sites: The sites the signed-in person saved on arche.directory, newest first.; Inputs: none

Measuring a built page for check_page

check_page compares numbers, so it needs the built page measured the way Arche measures the archive. The measuring script is public at https://arche.directory/check.js. It reads the page’s computed styles and the animations it runs, scrolls once from top to bottom and returns a JSON object. It sends nothing anywhere.

Open the built page in a browser at a 1440×900 viewport: Playwright, Chrome DevTools or a browser tool your agent has.
Evaluate the contents of check.js in the page and wait for it to resolve (up to about ten seconds on a long page).
Pass the returned object, unchanged, as measured, with the reference site as slug.

The answer lists where the two differ, with both numbers: the headline at 28 px against 100 px, no scroll reveal against a 900 ms one. Differences are not errors: the reference is one well-made answer, not a target. Use the numbers to spot weak craft (a flat type scale, low contrast, cramped spacing) and keep the differences that are your own choices.

Using a site as a reference

The sites in Arche belong to their owners. Use them the way designers use references: a starting point to learn from (type scale, spacing, density, colour temperature, how much moves), mixed with your own ideas and brand, and built with your own name, copy and images, so the result is recognisably yours. Don’t copy a site’s text, logo, images or exact layout. Measured values (colours, sizes, durations, easing) are facts about a page, and the motion code is written by Arche from those values. Site owners can ask for a correction or removal at arche.directory/removal.

Your data

To connect, you sign in with Google; Arche keeps your name, email address and Google account ID, the sites you save, and whether and when you first connected a client. Arche counts which tools are called and, when a client connects, its name (for example “Claude”). It doesn’t store your queries, the pages you measure or anything else from your conversation. Details, retention and deletion are in the privacy policy.

Limits and price

60 requests a minute per account. Arche MCP costs $7 a month or $40 a year and is free during early access; everyone who connected gets an email before it becomes paid, and nothing is charged unless you subscribe. See pricing.

Troubleshooting

The client can’t connect or keeps asking to sign in

Check the address is exactly https://arche.directory/mcp and that the client supports remote MCP servers with OAuth. Sign-in keeps working while the client uses it; after about a month without use the client asks you to sign in again. Signing out of arche.directory doesn’t disconnect a client.

“No site … in the archive”

Tools take a slug, a name or a domain (“linear”, “Linear”, “linear.app”). Search with search_sites first, or look the site up on arche.directory: the slug is the last part of its address.

A personal portfolio isn’t found

Personal portfolios are listed on arche.directory for people to browse, but not shared with agents unless their owner asked for it.

“… has no measured system to compare with yet”

Not every site has been measured yet. Pick another reference from search_sites; get_site says which data a site has.

check_page says the input isn’t check.js output

Pass the whole object the script returned, unchanged. It starts with "arche": 1 and has a styles list. Run it again after the page has fully loaded if it came back empty.

“Too many requests”

Each account can make 60 requests a minute, shared by all its connected clients. Wait a minute and try again.

How do I disconnect?

Remove the connector in your client. Deleting your Arche account on the Saved sites page also revokes every connection at once.

Support

Write to hello@arche.directory for help, bugs or security issues. For how to use the archive with any AI, see website design references for AI.

Arche MCP documentation. How the server works, what each tool does, how sign-in and your data are handled, and what to do when something goes wrong.