bol-mcp

MCP server that exposes bol.com's "Shophulp" AI shopping assistant as one tool, ask_bol. Sibling of mediamarkt-mcp.

ChatGPT calling ask_bol with BOL_UI=widget: one MCP Apps carousel row per Shophulp recommendation group, the model works from the text and structured data.

How it works

bol's Shophulp page POSTs an OpenAI-style streaming request to https://www.bol.com/streaming/api/v1/chat/completions (model gemini-3.5-flash, stream: true, user = a chat/thread UUID). No login or cookies are needed; a browser User-Agent and Accept: text/event-stream suffice. The SSE stream carries normal chat.completion.chunk deltas plus t800.tool_outputs chunks with grouped product recommendations (title, summary, products with price, seller, rating, image, url), a follow-up question and suggestion chips. This server flattens that into text + structuredContent.

Unofficial endpoint: it can change or be blocked at any time (bol.com sits behind Akamai Bot Manager). See bol's privacy policy.

Related MCP server: MCP-project

Install

npm install
npm run check   # live smoke test against bol.com

Run

node index.mjs                 # text mode (default)
BOL_UI=widget node index.mjs   # widget mode (MCP Apps carousel)
PORT=3001 node index.mjs       # use another HTTP port
node index.mjs --stdio        # stdio transport instead

Widget mode is opt-in because its UI domain must be configured for the deployment. For a remote host, set BOL_MCP_WIDGET_DOMAIN to that host's reachable HTTPS origin, for example https://mcp.example.com (origin only, without /mcp). This is separate from BOL_MCP_PUBLIC_URL, which is the MCP endpoint and includes /mcp. Then set BOL_UI=widget; otherwise leave it unset for text mode.

Connect

codex mcp add bol --url http://localhost:3000

For ChatGPT, expose the HTTP server publicly (e.g. cloudflared tunnel --url http://localhost:3000) and add the URL under Settings → Apps in developer mode. Run with BOL_UI=widget for the carousel.

Tool

ask_bol({ question }) — any language; Shophulp answers in the language of the question, product names stay Dutch. Returns content[0].text (assistant text, or the group summaries plus follow-up question when Shophulp answers with groups only) and structuredContent with products[] (id, name, group, price, currency, seller, rating, reviewCount, image, url), groups[] (title, summary), followUp and suggestions[].