How do I expose my product catalog to ChatGPT and Claude via MCP? (2026)

The short answer

Host a remote MCP server: one HTTPS endpoint that exposes your catalog as agent-callable tools — search products, get full specs, compare models, request a quote. Both Claude and ChatGPT support remote MCP servers, so customers install nothing: a buyer (or their company's agent) adds your endpoint once and then queries live catalog data mid-conversation instead of relying on whatever the model half-remembers from training. For a typical catalog the build is a thin API layer over data you already have — the real work is making the underlying product data complete and consistent enough to be worth querying.

The six steps

Consolidate the catalog into queryable records. One structured record per product: name, model identifiers, specs as typed fields (not PDF prose), price or price-on-request, availability, compatibility relationships. Gaps here become hallucinations later — an agent that can't retrieve duty_cycle will guess it.
Define a small tool set. Four tools cover most catalogs: search_catalog(query, filters), get_product(id) returning the full record, compare_products(ids), and for B2B request_quote(product, contact). Resist the temptation to expose twenty endpoints — agents pick tools by description, and a tight set with precise descriptions outperforms a sprawling one.
Serve it over Streamable HTTP. Remote MCP means one public HTTPS route handling the MCP JSON-RPC methods (initialize, tools/list, tools/call). SDKs exist for TypeScript and Python; a catalog MCP server is typically a few hundred lines plus your data access.
Connect it to the agents. Claude: claude mcp add --transport http yourbrand https://mcp.yourbrand.com/mcp, or add it as a custom connector in Claude's settings. ChatGPT: add the server under connectors/developer mode, or reach it through ACP-partnered platforms as those roll out. Test with a real buying conversation, not just a tools list.
Make it discoverable. Publish /.well-known/mcp.json on your domain pointing at the endpoint, reference it in llms.txt, and if you maintain an agent catalog file (ai-catalog.json), list the MCP endpoint there with your other agent-facing surfaces. Discovery standards are young, but the well-known pattern is what early agent crawlers actually fetch.
Log and measure. Log every tools/call with the query text (redact contact fields), watch which questions buyers actually ask, and feed unanswerable queries back into catalog enrichment. An MCP server is also a market-research instrument — the queries are verbatim buyer intent.

Design details that matter in practice

Return structured JSON with a stable schema, plus a short human-readable summary. Agents compose answers from both; pure prose is hard to filter, pure JSON without field descriptions gets misread.
Quote-based pricing: return price_type: on_request explicitly. Leaving price absent or zero produces agents quoting $0. Model the commercial reality instead of omitting it.
Rate-limit per client and require no auth for read tools. Frictionless reads maximize adoption; put auth only on write-shaped tools like quote submission, and validate/sanitize their inputs — tool arguments are untrusted user input.
Version the tool descriptions. Agents cache capability lists; when you change semantics, change the tool name or description explicitly rather than silently altering behavior.
Verifiable sourcing strengthens your answers. If records carry signed spec credentials or per-field datasheet URLs, say so in the tool output — it gives an agent a concrete reason to prefer your figure over a marketplace scrape when the two disagree.

A worked example

We run this pattern in production: a hosted catalog MCP endpoint serving multiple vendors' office-equipment records — search_catalog, get_product, compare_products, request_quote — added to Claude with one command. In our 2026 evaluation battery of hard long-tail B2B queries (roughly 30 per surface), external assistants without catalog access returned the wrong product on 54–67% of them, depending on the surface; the same assistant with MCP catalog access stopped fabricating because it could retrieve instead of recall. That is the core value: MCP converts "the model's memory of your products" into "your products".

FAQ

Does ChatGPT support MCP servers directly? Yes — remote MCP servers can be added to ChatGPT (connector/developer settings). OpenAI's broader commerce stack, the Agentic Commerce Protocol, is feed-and-endpoint first but added MCP integration in its 2026-04-17 release, so the two tracks are converging. Practical read: MCP gets you queryable in Claude, ChatGPT, and most agent frameworks; ACP enrollment is the separate track for in-chat checkout.

Do I need MCP if I already have good JSON-LD on my product pages? They serve different retrieval modes. JSON-LD is passive: crawlers ingest it on their schedule, and it feeds citations in answer engines. MCP is active: the agent queries live data at conversation time — current stock, current price, filtered search. B2B configurational questions ("which model supports X with accessory Y?") are effectively only answerable through the active path.

Is exposing my catalog to agents a competitive risk? Your product data is already public on your website — scraped, embedded in training data, and paraphrased with errors. An MCP endpoint doesn't reveal more; it makes the accurate version the easiest one to retrieve. Keep genuinely sensitive data (contract pricing, inventory by warehouse) out of the read tools, exactly as you keep it off your public site.

How long does an MCP catalog server take to build? With clean product data: days, not months — the protocol handling is boilerplate and SDKs are mature. The honest worst case is the other path: with specs scattered across PDFs and legacy systems, consolidation into queryable records has taken us weeks per catalog — and that work is needed for every other agent surface anyway (feeds, JSON-LD, ACP/UCP). Budget the effort there, not in the protocol layer.

How do buyers find my MCP server? Today: you tell them — a developer/agent page on your site, the well-known discovery files, and direct onboarding of key accounts ("add our catalog to your Claude"). Agent-side automatic discovery is emerging (registries, ai-catalog.json crawling), but in 2026 distribution is still push, then protocols. The measurable adoption signal is initialize calls from client IDs you didn't onboard by hand.