Ottom / MCP server
An ecommerce MCP server for AI assistants
Ottom’s MCP server is a remote, read only Model Context Protocol server that gives AI assistants data on how ready online stores are for AI shopping agents. An assistant connected to it can run a light check of any public store page, read Ottom’s published agentic commerce benchmark for a store category, and list what the check looks at. It answers at ottom.io/mcp, and keys come with every paid Ottom plan.
| Address | ottom.io/mcp |
|---|---|
| Transport | Streamable HTTP |
| Tools | 3 |
| Calls per key | 50 a day |
| Checks per store | 3 a day, across every key |
| Access | Every paid plan |
What data you can get
The server has 3 tools. Your assistant chooses which to call from your question. Each answer says where its data comes from.
check_store_readiness
Check how ready a store is for AI agents. It takes url, and returns:
Read now from the store’s public pages
- Whether the page was read, and if not, why: robots.txt asked Ottom not to, the address was refused, no page came back, or the store is behind a storefront password.
- What the store publishes for agents beside the page: its robots.txt and whether it lets agents read this page, an llms.txt file, a product feed and how many products it lists, and a Universal Commerce Protocol manifest and whether it is valid.
- What the page’s own markup says: how many JSON-LD blocks it has, whether schema.org Product markup is there, whether that markup carries a price, and whether the price is in the text an agent reads.
- Whether a bot wall answered instead of the page.
- Each finding with its stall code, its severity, the stage of a purchase it affects, a plain explanation, a fix written to forward to a developer, and a link to read more.
- What the light check did not look at, named, so an absence never reads as a pass.
Only when the store’s owner publishes it
- The store’s own Ottom result: PASS, SLOW, STALLED or NOT CHECKED RECENTLY, its score out of 100, when the full scan behind it ran, and its public result page. Only when the store’s owner publishes it with the Ottom badge; empty for every other store.
About Ottom’s own checks
- Where to start a full scan, in which an AI agent tries to find, compare and buy.
get_benchmark
Get the public benchmark for a store category. It takes category, and returns:
Aggregate, never a named store
- How far AI agents get at each of the four stages, discover, evaluate, compare and buy: the share of stores that passed and the median stage score.
- The share of stores scanned where an AI agent reached the payment step.
- The spread of Ottom scores: the median and the middle half, out of 100.
- The stalls that stop agents most often, each with its stall code, its stage and the share of stores it was found on.
- The benchmark in a few sentences, with a link to its page on ottom.io.
- The licence, CC BY 4.0, and the methodology, so any figure can be cited.
Categories: all, apparel, beauty, electronics, food-and-drink, health-and-wellness, home, jewellery-and-accessories, outdoor-and-sports. A category answers once enough of its stores have been measured, and the same figures are on the benchmark pages.
list_checks
List what the light check looks at. It takes no arguments.
About Ottom’s own checks
- Every check the light check runs, by name, with what it looks for.
- How much one check reads, and the daily limits per key and per store.
- What the light check never does.
The checks it lists:
- robots.txt. Whether the store’s robots.txt lets AI agents read the page, and any crawl delay.
- llms.txt. Whether the store publishes an llms.txt file for AI agents.
- Product feed. Whether a machine readable product feed is there, and how many products it lists.
- UCP manifest. Whether the store publishes a Universal Commerce Protocol manifest, and whether it is valid.
- Structured data. schema.org Product markup on the page: price, availability and a GTIN or MPN.
- What an agent can read. Whether the price appears in the text an agent reads, not only in images or scripts.
- Bot protection. Whether a bot wall answered instead of the page.
- Returns and shipping. Whether the returns and shipping policies the page links to can be read.
A finding’s stall code is the same one a full scan reports. Every code is described in the stall library and published as data in taxonomy.json. How the benchmark figures are made is on the methodology page.
Questions you can ask your assistant
- “Is this product page ready for AI shopping agents? yourstore.com/products/linen-shirt”
Calls
check_store_readinessA light check of that page, run now: whether robots.txt lets agents read it, whether the price is in the markup and in the text an agent reads, what llms.txt, the product feed and the UCP manifest say, and each finding with a fix to forward to your developer. The assistant also says what was not checked.
- “How far do AI agents get on apparel stores, and where do they stall?”
Calls
get_benchmarkThe published apparel benchmark: the pass rate at each stage, the share of stores where an agent reached payment, the score spread, and the stalls found on the most stores, with the window the figures cover and the methodology behind them.
- “Compare my product page with the benchmark for my category.”
Calls
check_store_readinessandget_benchmarkBoth answers side by side: what the light check found on your page, and the stalls that are most common in your category, so you can see whether yours is one of them.
- “Across all stores, what most often stops an AI agent from buying?”
Calls
get_benchmarkThe pooled benchmark across every store Ottom has scanned: the most widespread stalls by share of stores, each with its stall code and the stage it hits, ready to cite with the licence and link that come with it.
- “What exactly does Ottom check on a store, and what does it never do?”
Calls
list_checksThe list of checks the light check runs, how much one check reads, the daily limits, and what it never does: open a browser, call an AI model, add to cart or pay.
Who it is for
- Merchants
Ask about your own product pages while you work on them, and get the fix in words you can forward to your developer.
- Agencies
Check a client’s pages from the assistant you already use, and set them against the benchmark for the client’s category.
- Developers
Check a page before and after a change to its markup, feed or robots.txt, without leaving the editor.
- Researchers
Read the agentic commerce benchmark by category as structured data, under an open licence, with the methodology one link away.
- Journalists
Quote a benchmark figure with its source, its window and its licence attached, rather than a number from memory.
Access
The MCP server is included with every paid Ottom plan. Make a key in your account, add the server to your assistant with the key as a bearer token, and ask. Listing the tools needs no key. Compare the plans, or read the MCP server reference for the address, setup for your client, limits and error codes.
What it never returns
- Anything from an Ottom account: no reports, no scans you ran, no findings from a full scan.
- A store’s own Ottom result, unless its owner publishes it with the Ottom badge.
- A named store in a benchmark. Benchmarks are aggregate and anonymised.
- A visit to a store that has opted out of Ottom’s agent.
- A browser, an AI model, a cart, a checkout or a payment. The light check reads one page and a few files.
- A visit from your machine. Every check comes from Ottom’s published addresses and user agent.
How Ottom identifies itself when it visits a store, and how a store opts out, is on our agent page.
Common questions
What is an MCP server?
A server that speaks the Model Context Protocol, an open standard for connecting AI assistants to outside tools and data. An assistant that supports MCP lists the server’s tools, decides when one would help answer your question, calls it, and reads the answer.
What data can an AI assistant get from Ottom’s MCP server?
Three kinds. A light check of any public store page, run when asked: robots.txt, llms.txt, the product feed, the UCP manifest, structured data and whether the price is in text an agent can read, with findings and fixes. Ottom’s published agentic commerce benchmark for a store category or for every store. And the list of what the light check looks at.
Which plans include the MCP server?
Every paid Ottom plan. Keys are made on the MCP keys page in your account, and each key allows 50 calls a day. Listing the tools needs no key, so an assistant can see them before you add one.
Which AI assistants can use it?
Any MCP client that connects to a remote server over Streamable HTTP and lets you send a key in the Authorization header. Claude Code, Cursor, VS Code and Gemini CLI document exactly that setup. The reference has the address and a configuration to copy.
Can it show me a competitor’s Ottom score?
No. A store’s own Ottom result is returned only when its owner publishes it with the Ottom badge. The light check reads any public page that has not opted out, the same way a page speed tool does, and says only what that page publishes to anyone who asks.
Can I cite the benchmark figures it returns?
Yes. The benchmarks are published under CC BY 4.0, and every answer carries the licence, the window the figures cover, a link to the benchmark page and the methodology. The figures are aggregate and never name a store.
Does checking a store visit it?
Yes, once per check, from Ottom’s published addresses with its published user agent, never from your machine. A store that has opted out is never visited, robots.txt is read first, and any one store is checked at most 3 times a day however many people ask.
How is this different from a full Ottom scan?
The light check reads one page and a few files. A full scan sends an AI agent through discovery, comparison and checkout on the store and reports where it stalls. Full scans are run from ottom.io, and their results stay in your account.