AI integration
TheSeeker Reference MCP server
AI tools such as Claude, Cursor and Codex search the UI design references collected in TheSeeker Reference directly, and read as much of their analysis, context and images as they need.
https://ref.theseeker.io/mcpMCP (Model Context Protocol) is a standard that connects AI tools to outside data. Register the server once, and during a conversation the agent can look through tags, find references that match, check their images and answer with links to the originals.
The same data, presented to people and to AI agents in the way that suits each.
| Aspect | Web gallery | MCP |
|---|---|---|
| Audience | People browsing references | AI agents connected through an MCP client |
| Access | Browse and search in a browser | Tool calls the agent makes during a conversation |
| Address | https://ref.theseeker.io | https://ref.theseeker.io/mcp |
| Output | Card grids and detail pages | JSON results and image content blocks |
- An AI client that supports MCP. Connection steps for each client follow below.
- A network that can reach
ref.theseeker.io - No account or API key. Connect right away, with no sign-up or token.
Pick the client you use and register the server as described. In every client, name the server ref.
Run the command in a terminal.
Check the tool list in Claude Code.
claude mcp add --transport http ref https://ref.theseeker.io/mcpGo to Settings → Connectors.
Choose Add custom connector.
Enter ref as the Name and the address below as the URL, then press Add.
https://ref.theseeker.io/mcpAdd an entry under mcpServers in the config file, then reload Cursor.
~/.cursor/mcp.json{
"mcpServers": {
"ref": {
"url": "https://ref.theseeker.io/mcp"
}
}
}Add an entry under servers in the workspace MCP config file.
.vscode/mcp.json{
"servers": {
"ref": {
"type": "http",
"url": "https://ref.theseeker.io/mcp"
}
}
}Add an MCP server block to the Codex config file.
~/.codex/config.toml[mcp_servers.ref]
url = "https://ref.theseeker.io/mcp"Register the remote server under mcp in the config file. No authentication setup is needed.
opencode.json{
"mcp": {
"ref": {
"type": "remote",
"url": "https://ref.theseeker.io/mcp",
"enabled": true,
"oauth": false
}
}
}Add the connection under mcpServers in the MCP config file.
~/.codeium/windsurf/mcp_config.json{
"mcpServers": {
"ref": {
"serverUrl": "https://ref.theseeker.io/mcp"
}
}
}Add the connection under mcpServers in the Gemini CLI settings file.
~/.gemini/settings.json{
"mcpServers": {
"ref": {
"httpUrl": "https://ref.theseeker.io/mcp"
}
}
}Send the initialize request from a terminal.
See the tool list and tool call requests in the other examples.
curl -sS https://ref.theseeker.io/mcp -H 'content-type: application/json' -H 'accept: application/json, text/event-stream' -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"curl","version":"1.0.0"}}}'The full request format and curl examples are in the endpoint spec of the tool reference.
Agents usually find references in five steps. The more each result narrows the arguments of the next call, the faster and more accurate the answers get.
Check tags and categories
list_tagsExplores public reference counts by category, and the tags. Narrow the tags by category or search term.
Search for matching references
search_referencesFinds public references by keyword, tag, category and designer. Each result carries a preview image and the original page URL.
Compare candidates on one sheet
get_contact_sheetPuts up to 16 candidate references on one image for comparison. Each cell shows its reference id.
Read only the sections you need
get_referenceReads only the parts of a reference you pick: analysis, media, original post, designer and similar work.
View images of the chosen references
get_reference_imagesShows the images and video posters of chosen references directly. An error on one image does not block the others.
get_contact_sheet or split them across two, and request original images in small or medium size only for the 3–5 you will actually use. One call returns up to 6 images.Every tool is read-only. Select a tool name for its arguments and response examples.
| Tool | What it does |
|---|---|
search_references | Finds public references by keyword, tag, category and designer. Each result carries a preview image and the original page URL. |
get_reference | Reads only the parts of a reference you pick: analysis, media, original post, designer and similar work. |
get_reference_images | Shows the images and video posters of chosen references directly. An error on one image does not block the others. |
get_contact_sheet | Puts up to 16 candidate references on one image for comparison. Each cell shows its reference id. |
list_tags | Explores public reference counts by category, and the tags. Narrow the tags by category or search term. |
list_authors | Shows the bio and reference count of every designer with public references. |
Once connected, try asking the agent something like this.
- Find dark-mode fintech dashboards in TheSeeker Reference and summarize the layout patterns they share.
- Gather a wide set of bento-grid landing pages, compare them on a contact sheet, pick 3 and look at their images, then suggest a hero section that fits our product.
- Find the 3 designers with the most references and compare their styles.
Check tags first
Use list_tags to see the tag slugs actually in use, then pass them in tags of search_references. A tag that does not exist returns no results.
Include only the sections you need
Put only the sections you need in include of get_reference. The available values are analysis, media, tweet, author, similar; the less you request, the lighter the response.
Look wide, open originals for only 3–5
Compare up to 16 candidates on one sheet with get_contact_sheet, then call get_reference_images in small or medium size only for the 3–5 references you will actually use. Original images take up a lot of conversation context.
Cite pageUrl
Include the pageUrl that comes with each result in your answer, so readers can open the original reference right away.
Search in Korean or English
Titles and summaries are in Korean, and English fields such as titleEn and summaryEn are indexed too. Both “대시보드” and “dashboard” find results.
Retry-After header; wait that long, then try again.Image limits
| Images per call | 4 by default, up to 6 |
|---|---|
| Per image | Up to 5 MiB |
| Per call in total | Up to 15 MiB |
| Image fetch timeout | 10 seconds |
Error codes
| Code | Layer | Meaning |
|---|---|---|
RATE_LIMITED | Tool result | Contact sheet build and queue capacity is full; retry shortly. |
NOT_FOUND | Tool result | The requested reference does not exist or is not public. |
VALIDATION | Tool result | A value passed to the tool is outside the allowed range or format. |
INTERNAL | Tool result | Something went wrong during the lookup. Try again shortly. |
400/-32600 | HTTP response | At most one get_contact_sheet call is allowed per batch. |
405 | HTTP response | HTTP method not allowed. Use POST for tool calls. |
406 | HTTP response | Include both application/json and text/event-stream in the Accept header. |
429/-32029 | HTTP response | Per-minute request limit exceeded. A batch counts once per tools/call (at least once). Retry after the time given in Retry-After. |
Remove the ref entry from your client settings. The server has no sessions or tokens to revoke.
- Claude Code:
claude mcp remove ref - Claude Desktop: Delete the ref connector under Settings → Connectors.
- Cursor: Delete the ref entry, then reload.
- VS Code: Delete the ref entry from servers.
- Codex: Delete the [mcp_servers.ref] block.
- opencode: Delete mcp.ref, or set enabled to false.
- Windsurf: Delete the ref entry from mcpServers.
- Gemini CLI: Delete the mcpServers.ref entry.
- curl: The server is stateless, so there is no connection to remove.