Integrate with your agent

Your agent integrates, you review the diff. With the key, this documentation and the CrossUp MCP, it verifies the key, writes the snippet and tests a recommendation and a signal. Handing it over is safe: the storefront key is public and scoped. Why, in Keys and security.

What your agent needs

What	What for	Where from
The `spk_…` key and the `connection_id`	Request recommendations and send test signals	The admin, under Settings → Keys for your storefront
This documentation	Know how to integrate	`llms.txt`: the whole site in one file, or each page as Markdown
The CrossUp MCP	Read the catalog and request recommendations from the agent	The MCP
The integrator skill	Follow this documentation’s path, with its checks	The skill

The prompt

Paste this into your agent, with your key and your connection:

Integrate CrossUp into this store's product page.

Storefront key: spk_…
connection_id: cn_…
Documentation: https://docs.crossup.ai/llms.txt

1. Verify the key with a GET /v1/catalog/items of any item.
2. Install @crossup/storefront and create a single instance.
3. Request a decision with surface "pdp", placementKey "pdp-related" and the
   item_id of the page's product. Display results[].product with the store's
   components; if there are no results, display nothing.
4. Report track.impression when the recommendation enters the screen and track.click when a result is opened.
5. Show me the decision_id of a test decision and the accepted signal_id.

Don't make up values the API doesn't return. An absent field is unknown, not zero.

What an agent can do today, and what it can’t

With the storefront key
Read the canonical catalog	Yes
Request recommendations for any moment	Yes
Send signals	Yes, through the API; the MCP doesn’t expose them yet
Enable or configure moments	No: that’s the admin’s
Read analytics	No: that’s the admin’s
Create a connector	No: it needs a connector key

The key has three permissions and a quota of 60 requests per minute: an agent can’t do anything the key doesn’t allow.

What you will see

A diff with the instance, the request and the two signals. In the agent’s answer, a decision_id (the recommendation’s id) and a signal_id with accepted: the same evidence you’d ask a person for.

If you don’t see it

What happens	What to do
The agent makes up an endpoint or a field	Give it `llms.txt` or the docs MCP
It shows `0` where there’s no price	Paste the canonical model rule: an absent field is not a zero
It retries in a loop on a `429`	Paste Quota and cache
It asks for `store_id`	It doesn’t exist: the store comes from the key

The MCP

The tools an agent can call.

The skill

The integration path, as a file for your agent.

​What your agent needs

​The prompt

​What an agent can do today, and what it can’t

​What you will see

​If you don’t see it