Connect Your AI Coding Agent to Your Atelier

claude mcp add --transport http yeja-atelier https://yeja-ai.com/mcp

Then start a chat and give the agent your profile link and key.

Do not add --header "Authorization: …". A key in the settings ties the server to that one account in every chat. Any Authorization header also turns off OAuth fallback in Claude Code, and the header is read once when the process starts, so changing it later means restarting Claude Code. To sign in instead of using a key, add https://yeja-ai.com/mcp/oauth, then run

/mcp, pick the server and choose Authenticate.

Add it as a custom connector: open Connectors in Claude's settings, press +, choose Add custom connector, paste https://yeja-ai.com/mcp and press Add. Then start a chat and give Claude your profile link and key. The same connector is available in Claude Desktop and on claude.ai, and it works for any profile.

If an older Yeja connector is signed in, it can only ever work on the account it signed in as. Check the account on the dashboard's Connected Agents list, and remove the connector if it is not yours.

Custom connectors are available on every Claude plan; the Free plan allows one. On Team and Enterprise, an organization Owner adds the connector first and members then press Connect.

Advanced: A Key In The Settings

Claude Desktop's config file cannot send a custom header to a remote server, so a key needs the mcp-remote bridge:

{
  "mcpServers": {
    "yeja-atelier": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote", "https://yeja-ai.com/mcp",
        "--header", "Authorization: Bearer yeja_atl_YOUR_KEY_HERE"
      ]
    }
  }
}

{
  "mcpServers": {
    "yeja-atelier": {
      "url": "https://yeja-ai.com/mcp"
    }
  }
}

In ~/.cursor/mcp.json, or .cursor/mcp.json inside a project. Remote entries need no type field. Then give the agent your profile link and key in the chat.

Advanced: A Key In The Settings

{
  "mcpServers": {
    "yeja-atelier": {
      "url": "https://yeja-ai.com/mcp",
      "headers": { "Authorization": "Bearer yeja_atl_YOUR_KEY_HERE" }
    }
  }
}

The dashboard's Developer Access card also has a one-click install link for Cursor, which is easier than editing this file.

{
  "servers": {
    "yeja-atelier": {
      "type": "http",
      "url": "https://yeja-ai.com/mcp"
    }
  }
}

In .vscode/mcp.json, or in your user configuration (MCP: Open User Configuration). Note this client uses servers, not mcpServers. Then give the agent your profile link and key in the chat.

Advanced: A Key In The Settings

{
  "inputs": [
    {
      "type": "promptString",
      "id": "yeja-key",
      "description": "Yeja Atelier API Key",
      "password": true
    }
  ],
  "servers": {
    "yeja-atelier": {
      "type": "http",
      "url": "https://yeja-ai.com/mcp",
      "headers": { "Authorization": "Bearer YEJA_KEY_PLACEHOLDER" }
    }
  }
}

Replace YEJA_KEY_PLACEHOLDER with an input reference so VS Code prompts for the key rather than storing it in the file: a dollar sign, then

{input:yeja-key}. The dashboard's Developer Access card also has a one-click install link for VS Code.

codex mcp add yeja-atelier --url https://yeja-ai.com/mcp

Then give the agent your profile link and key in the chat.

Advanced: A Key In The Settings

codex mcp add yeja-atelier --url https://yeja-ai.com/mcp \
  --bearer-token-env-var YEJA_API_KEY

Codex takes the NAME of an environment variable, not the key itself. Set

YEJA_API_KEY in your shell to the key, and pass the name here. Pasting the key where the name belongs fails in a way that does not say so. A bearer token takes precedence over sign-in when both are set. The file form, in

~/.codex/config.toml:

[mcp_servers.yeja-atelier]
url = "https://yeja-ai.com/mcp"
bearer_token_env_var = "YEJA_API_KEY"

{
  "mcpServers": {
    "yeja-atelier": {
      "serverUrl": "https://yeja-ai.com/mcp"
    }
  }
}

Use serverUrl, not url. serverUrl works in both clients, while some versions ignore a plain url for a remote server, and an ignored entry looks exactly like a server that is not responding.

Antigravity reads ~/.gemini/config/mcp_config.json. Windsurf is now Devin Desktop, which reads ~/.config/devin/mcp_config.json. Then give the agent your profile link and key in the chat.

Advanced: A Key In The Settings

Add "headers": { "Authorization": "Bearer yeja_atl_YOUR_KEY_HERE" } beside

serverUrl.

{
  "context_servers": {
    "yeja-atelier": {
      "url": "https://yeja-ai.com/mcp"
    }
  }
}

Zed calls them context_servers, not mcpServers. Open the file with zed: open settings file. Then give the agent your profile link and key in the chat.

Advanced: A Key In The Settings

Add "headers": { "Authorization": "Bearer yeja_atl_YOUR_KEY_HERE" } beside

url.

{
  "mcpServers": {
    "yeja-atelier": {
      "type": "streamableHttp",
      "url": "https://yeja-ai.com/mcp"
    }
  }
}

Then give the agent your profile link and key in the chat.

streamableHttp is case-sensitive. Writing streamable-http, or leaving

type out, makes Cline fall back to SSE, and this server answers SSE with a 405 that does not explain itself.

curl -s -X POST https://yeja-ai.com/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

A list of tools coming back means the server is reachable. The agent then connects a profile in the chat; until it does, every tool except the connection tools answers not_connected. A 401 means a key in your client's settings is wrong or revoked.

A snapshot is taken automatically before every publish. restore_snapshot rolls your live page back to one, takes a snapshot of the current state first so the restore can itself be undone, and clears any pending draft. A good agent confirms with you before calling it.

For a Freeform page, rollback_freeform re-publishes an earlier version through the full publish gate. An artwork you have deleted since does not come back.

Staged work you no longer want: discard_staged_rows and

discard_freeform_draft. Your live page is untouched by both.

To start the theme over, reset_atelier_theme puts your Atelier back on the default Yeja theme: palette, type, corner radius, background and custom CSS are cleared, and an installed or agent-built Freeform page is taken down and its draft discarded. Your works, statement, timeline, sections and widgets stay. Called without confirm: true it only lists what would change. It snapshots first, so restore_snapshot undoes it, except a discarded Freeform draft.

To take only the Freeform page down, unpublish_freeform puts visitors back on the classic Atelier and keeps the draft and every version; rollback_freeform brings it back. To take the whole Atelier offline, unpublish_atelier does what Unpublish Atelier in the dashboard does: nothing is deleted, and only you can publish it again. Both are dry runs until called with confirm: true.

• Pinning a work with no published image. It would render as a blank card. Publish the image in My Studio first.
• Pinning a trashed work.
• A partial reorder. reorder_pins needs every pin named, in order.
• Deleting a product with orders. Archive it instead.

remove_timeline_entry cannot be undone; remove_section can, with

restore_section.

A good agent starts with six questions: what the page is for, who lands on it, what a visitor should do, how it should feel, what must appear, and which sites you want it to resemble. About two minutes. An agent can see your artworks but none of that, and that is what decides the layout.

draft_from_reference refuses to build until the six are answered. Say plainly that you want to skip the questions and the agent can proceed; that it did is recorded. replicate_site does not wait: it fills only the answers you have not given with provisional ones, marked in your dashboard as the agent's and not yours, so check and correct them. Your answers are saved, so the interview happens once; say "start over" to rethink the page. None of this is public.

Ask for everything and a good agent fills every surface, not just the banner and a grid:

A section you fill but never switch on stays invisible, so an agent should turn on whatever it populates. And it should ask you for anything it does not have (dates, collaborators, a statement) rather than writing something plausible. A made-up exhibition history is worse than an empty one.

Your agent is instructed by the server never to invent a testimonial, a price, a class date, a venue or a credential. Without the real thing from you, it leaves the section out and says so. Still read anything an agent writes about facts before you keep it.

Selling through Yeja is coming soon. The Yeja Wallet is under development; once it is live, you will be able to sell your work and your themes directly through Yeja. Until then, point visitors at wherever you already sell.

Products (a class, a course, a workshop seat, a book, a print, a download) can be created, priced and shown on your page. Put an outside shop link in

external_url. New products start as drafts, invisible until set to available.

Atelier Themes: package_current_freeform_as_theme turns your published Freeform design into a listing with your artworks, bio and products replaced by marked placeholders. It starts as a draft; show_theme_listing puts it on Art Board, where other artists can browse it. Visitors who want it can reach you through your Atelier's "Request a Theme" link.

install_theme writes a purchased theme into the buyer's own Freeform draft, never their live page.

• inspect_atelier_render screenshots your page at desktop, tablet and mobile, in light or dark. It shows the published page by default, whether that is your standard Atelier or a published Freeform page, and says which in live_mode. Pass preview: true to screenshot your unpublished standard draft, or freeform: true for your Freeform draft.
• inspect_layout_metrics finds sideways scrolling, failed images and heading skips. inspect_accessibility measures contrast on the colors that actually painted; run it in both themes.
• create_preview_link makes a link to your staged standard-Atelier draft, rendered exactly as visitors would see it, with a banner across the top. It refuses when nothing is staged, and it does not show staged collections (compare_public_and_draft lists those). It lasts 24 hours by default, 7 days at most, and is revoked the moment you publish or discard that draft.
• atelier_audit ranks what to fix. When your live page is a Freeform page, it also checks that page for unfilled placeholder text.
• analyze_artworks reports pixel size, 300dpi print size and dominant color for your images; videos are listed as skipped, not as failures. The centimeters on an artwork are what you entered; they are never derived from the file.
• get_atelier_views reports visits, sources and countries. It says "no data yet" rather than zero for pages older than 2026-08-26, has no per-artwork numbers, and stores no raw IP address or browser string.

inspect_seo reports what a crawler and a link preview receive, including a missing share image and any drift between your saved SEO fields and the page.

propose_seo suggests values from your own bio, skills and pinned works and writes nothing; an agent stages them with propose_atelier, and nothing reaches a crawler until you press Save SEO yourself.

There is no language model behind this server; your agent writes, and the server checks. check_copy says whether a line survives the save unchanged: a headline is at most 80 characters of plain text, a bio at most 600 characters with simple formatting. review_alt_text reports what each work announces to a screen reader. That text is the artwork's title, so a weak title should be raised with you, not renamed.

Deep-color field, logo pill with a menu button, a staggered wall of tiles. Classed supported (97% predicted).

• First draft 17%. Plain white instead of the deep field, no logo pill, tiles overlapping on mobile.
• Round 2: 50%, held down by one real defect: platform text at 1.12:1 contrast on the colored field. Fixed.
• Round 3: 94% (0.938 desktop, 0.944 mobile). Column stagger within a few pixels of the reference, 11 working links, the menu opens, nothing dead.
• Touch-ups left: the catalog has no stencil typeface, so a monospace stands in, and the wordmark is one line where the reference stacks two. The artist had 5 works, so they repeat across 24 tiles.

Takeaway: a well-supported page type gets close in three rounds. Upload enough works to fill a big grid; the tool will not invent more.

Wordmark and navigation, a large title, an asymmetric grid of narrow and wide project tiles. Classed supported with fallbacks (83% predicted).

• The loop reached 100% desktop, 94% mobile in 2 rounds. Grid, captions and title matched; the only mark lost was a slightly larger mobile title.
• The first draft now scores 94% on its own, after the 2026-09-30 generator improvement. Before it, the first draft scored 0.
• The cost of making every link work: navigation dropped to 2 links where the reference has 6, because pages the artist does not have were left out rather than linked nowhere. That, plus a plain-list mobile menu, gives a final 94% on both.

Takeaway: grids, spacing and type carry over best. Count the navigation links yourself and ask for any you want back.

A search bar, a filter column, category chips and rows of events with share and save icons. Classed supported with fallbacks (62% predicted); booking cannot be rebuilt.

• First draft 56% (0.625 desktop, 0.556 mobile).
• After 4 rounds: 100% desktop, 94% mobile. Header, breadcrumb, filter column, chips and event rows all in place.
• Behavior repair: 94% on both. Search, filters, chips and share/save are static look-alikes: they look right and do nothing, because a published Atelier runs no custom script. "Updates" and "Help Center" were dropped, and on mobile the icons sit below the image instead of over it.

Takeaway: interactive features come across as appearance only. For anything a visitor must do, like buy a ticket or get in touch, use a real link.

classify_reference_site recognizes eight kinds of page: exhibition or gallery, photography portfolio, architecture studio, editorial lookbook, design agency, artist portfolio, event landing page, and professional profile. Pages built around a WebGL scene, or behind a password, are unsupported. Forms, sign-in, booking, carts, file uploads and cross-origin embeds come back with a named fallback.

• Anything that needs script: a reference's own forms are replaced by the platform contact form, which sends to your inquiries; search, filters, carts and booking become static; a self-advancing carousel and WebGL do not work.
• Fixed bars: position: fixed is not allowed, so a newsletter or cookie bar is dropped.
• Typefaces map to the six in the catalog.
• Some images are missed: slides in a clipped carousel, and some lazy-loaded hero photographs.
• Brand marks can land in the wrong place, and a reference's own poor contrast can be copied faithfully.
• Some sites refuse to be measured (a 403), and nothing can be built from them.

• Navigation: count the links against the reference.
• Placeholders: replace "Artist Name" and any marked placeholder text.
• Enough works to fill the grid without repeats.
• The mobile menu: open it on a phone.
• Contrast in both themes.
• The hero image is actually there.
• Contact: every "Contact" or "Commissions" link points at the page's contact form.
• Then ask for atelier_audit and fix what it ranks first.

A clone copies a layout; it does not know about your Yeja widgets. Ask for them explicitly: "After the clone, add every widget I can support." The agent calls get_widget_inventory, which reports each one as present, empty, supportable or not supportable:

An agent should place every widget with real content and never fabricate one.

The rubric has nine criteria, each scored 0, 0.5 or 1: opening composition, navigation, section order, proportions, image treatment, type scale, whitespace rhythm, color, and on mobile, responsive behavior. A round's similarity is the lower of desktop and mobile. compare_round also counts dead links and menus that do not open, and will not call a round passed while any remain.

The structural checks (compare_to_reference,

check_replication_fidelity) measure section rhythm, grid shape, type scale, palette and navigation. A page can score well there and still look wrong, so if an agent quotes you a percentage, ask to see the screenshots beside the reference.

Almost everything that makes a page feel alive is CSS, and CSS is allowed: hover states, transitions, transforms, keyframe animation and blend modes. Menus, in-page anchors and disclosure panels work on published pages. Fifteen named effects can be attached to an element. Ten compile to CSS:

Five run one of the platform's own signed scripts, which a published page is allowed to load: cursor_follower, smooth_scroll, ambient_audio,

scroll_linked_columns and scroll_linked_columns_endless. Check

get_freeform_runtime_capabilities before using one: it says which are deployed.

A page may carry up to 30 effects, and up to 3 on any one element. Out-of-range values are refused rather than clamped, and every effect stops under reduced motion.

Both work on a published page. A contact_form renders a real form (name, email, message) that sends straight to your inquiries; the page allows that one destination and no other. A support block renders one link per amount, and the pledge completes in the signed-in Yeja app, once Stripe payouts are connected.

get_freeform_schema lists every node type, the nesting limits and what the stylesheet may contain. Then call get_freeform_runtime_capabilities before designing: the schema says what validates, the runtime report says what actually renders in this deployment, and they differ. Before handing over a preview, validate_freeform_draft checks a tree or the saved draft without writing anything, and flags embeds of artwork you do not own.

create_freeform_draft validates before storing. update_freeform_draft patches only what it is sent. patch_freeform_css adds a rule or two without rewriting the stylesheet.

A version holds a site, not a page: extra pages go under a __pages key and publish together. Components let one card be defined once and placed forty times, styled by one rule.

build_artwork_gallery writes a whole wall of your works: columns, embeds, grid CSS and mobile breakpoints, generated from the same numbers so they cannot disagree. It returns the tree by default. With apply: true it adds the gallery to your existing draft; only mode: "replace_page" replaces the whole page. A gallery card is an image, so video works are refused: place each one with a video node (list_artworks lists them as usable_video_ids).

motion: "scroll_linked_columns" makes alternating columns drift as the page scrolls, without capturing the wheel. scroll_linked_columns_endless loops the columns and does read the wheel, but only while the gallery fills most of the screen, and it gives the wheel back after about 900px. Neither runs under reduced motion.

Inter, Playfair Display, JetBrains Mono, Plus Jakarta Sans, Yuyu and Geist Pixel are self-hosted. Naming any other family is a rejection, not a silent fallback. Generic families (serif, sans-serif,

monospace, system-ui…) are always fine. Yuyu (handwriting) and Geist Pixel (pixel display) each have one weight, 400; use Geist Pixel for headings, not body text.

There is no font upload and no Google Fonts import; both are refused with

FONT_UPLOAD_UNAVAILABLE. Adding a family is a request to the Yeja team.

Use embed_artwork for a work a visitor should be able to open. It renders the real image server-side, and the platform's viewer opens it in place, over the page you built: zoom, drag to pan, previous and next across every embed, Escape to close. A detail panel shows the work's own title, artist, availability, medium, year and provenance, and leaves out anything you never filled in. It has no comments, share or report controls, because a published page cannot make requests.

It degrades to the image, never to a blank space. artworkId must be a work you own (list_artworks); any other id renders nothing at all. It is for images: a video in embed_artwork or an image node shows as a broken image, and validate_freeform_draft warns about it.

A generic embed (a third-party page in a frame) publishes as an empty space, because a published page cannot load outside content. Leave it out, or use a labelled static section instead. The other widget nodes work:

contact_form is a real form, and embed_product renders a card for one of your products (title, price or "Enquire", and a link to where you sell it) with no buy button.

preview_freeform renders the draft, reports whether it would pass the publish gate, and returns a fingerprint. Publishing without a fingerprint is refused, and any edit changes it, so an approval cannot carry over to a changed draft. get_go_live_link gives you the dashboard link where you publish.

A Freeform preview link keeps working until it expires, and always shows what the draft holds now. After the Freeform draft is published or discarded, the link shows your published page, with a banner saying so. Publishing or discarding your standard draft revokes every preview link. refresh_preview_link mints a new one. snapshot_round pins a link to one round of a clone, so it keeps showing that round after the draft moves on.

list_freeform_versions shows every version, get_freeform_history every publish, rollback and unpublish, and rollback_freeform puts an earlier version live again.

These are enforced server-side, not merely requested. A proposal that breaks them is rejected or corrected, and the reason is returned to the agent.

• Connect the profile the user named, before anything else. connect_profile with the profile exactly as the user gave it. Every other tool refuses until it succeeds, and a signed-in connection refuses any profile but its own. On identity_mismatch, stop and tell the user; never work on the signed-in account instead.
• portfolio_process_scale and portfolio_statement_scale are the image's share of the row, not a zoom level. Values near 100 make the image take the whole row and push the page past the viewport. Keep them between 40 and 70.
• An agent cannot introduce a new image URL. A page may only reference media already in your account. An outside image must first be imported with import_artwork_from_url, with your permission and its source recorded.
• portfolio_headline is at most 80 characters; the bio is at most 600 characters of text.
• portfolio_grid_columns is 2 to 5. At most 3 process images and 8 social links.
• portfolio_sections stores only the OFF flag. An absent key means the section is visible; set a key to false to hide it.
• portfolio_section_order sets the order, top to bottom, from works, exhibitions, statement, about, skills, collaborations, inquire, socialLinks, products, auctions, campaign, drop, support and custom. A section left out keeps its default place at the end; leaving it out never hides it.
• Never invent a fact about an artwork. No dimension, price, material, date, exhibition or provenance detail that was not supplied.