Commerce CensusContactRequest access

Home / Solutions / AI shopping agents

The commerce API for AI shopping agents

An agent that returns a price has done a lookup. An agent that says whether the price is good has done the job. The gap between those is a history the agent does not have and cannot build inside a conversation.

MCP, six toolsSurfaceprice_contextKey callNamespaced memoryState

AI shopping agents

Verdicts, not readings

price_context returns excellent, good, typical, poor or flat with the percentile and window behind it, so the agent can state its reasoning rather than assert a conclusion.

AI shopping agents

Memory across sessions

Preferences, constraints, rejections and past purchases persist per namespace. The next conversation starts from what is already known instead of re-interviewing the user.

AI shopping agents

Metered per tool

One prompt fans into dozens of calls with nobody approving each one. Each tool declares its cost so an agent can budget, and a cheap metadata lookup never bills like a live cross-seller compare.

Endpoints that do this

Where it lands in the API.

MCP @commercecensus/mcp

MCP server

Six tools, not thirty endpoints. A chatty server makes an agent burn context stitching fragments and gives it room to hallucinate the join.

→ reference
GET /v1/products/{id}/context

Price context

Is this actually a good price? Percentile against the product's own history at that seller, time-weighted, with low, median and high.

→ reference
POST /v1/memory/query

Recall

Semantic recall scoped to one namespace. Ask what matters for the decision in front of you rather than replaying an entire history into context.

→ reference

Questions

Common questions.

Why not just let the agent browse?

A browsing agent sees one page at a time, has no history, and cannot tell whether the number in front of it is high or low. It also cannot answer the question without spending a great deal of context to still get it wrong.