AI agents

Amazon data tools your AI agents can call safely

Give agents typed Amazon operations through REST, GraphQL, MCP, and an installable Skill—without handing them unrestricted browsing or a second authorization model.

Start free See the recipe

/api/amazon

One tool catalog, every agent interface

GET/api/amazon/searchGET/api/amazon/productGET/api/amazon/product/reviewsGET/api/amazon/product/offers

Agent tools need bounded inputs, predictable cost, freshness, and typed failures. Glade API exposes the same 17-operation catalog through every interface and returns marketplace content as untrusted data, never as instructions.

What teams build with it

These are bounded agent tools with explicit schemas and costs, not open-ended browser tasks.

Search before retrieval

Let the agent produce a bounded candidate set, then resolve only the products needed for the answer.

Endpoint chainsearch -> product

Price checker tool

Return current offer context with marketplace, condition, seller, delivery, and fetched time attached.

Endpoint chainproduct -> offers

Review evidence tool

Fetch an explicit review page and treat all returned customer text as untrusted source data.

Endpoint chainproduct -> reviews

Seller inspection tool

Resolve marketplace-scoped public seller facts without asking the model to infer legal identity.

Endpoint chainseller -> offers

Best-seller discovery

Start from a bounded category list, then resolve the product records an agent needs.

Endpoint chainbestsellers -> product

Identifier conversion tool

Expose deterministic ASIN and GTIN conversions instead of asking a model to guess a mapping.

Endpoint chaingtin-from-asin <-> asin-from-gtin

How bounded agent retrieval works

Four controls keep an agent request inside the same documented contract used by every other Glade client.

Marketplace and provider content is data. Tool descriptions, input schemas, and system policy remain the instruction boundary.

GET /api/amazon/search
Validate a typed tool call
The agent selects a named operation and documented inputs rather than constructing arbitrary provider requests.
allowlisted operation schema
GET /api/amazon/product
Route through the canonical catalog
REST, GraphQL, MCP, and the Skill resolve to the same operation, metering, cache, and authorization path.
one authorization model
GET /api/amazon/product/reviews
Return untrusted marketplace data
Free-form titles and reviews stay in response data and are never promoted into tool or system instructions.
injection-resistant boundary
GET /api/amazon/product/offers
Surface freshness and usage
The agent receives typed success or failure metadata, fetched time, cache state, and predictable unit cost.
cost and freshness visible

APIs ready for agent workflows

Every operation below is available through the same canonical catalog. Open one to inspect its precise inputs and response contract.

Compared with the usual setup

Open-ended browsing is difficult to authorize, meter, cache, and evaluate. A bounded tool catalog makes each of those controls explicit.

Capability	Glade API	Typical setup
Tool surface	17 named Amazon operations with documented, typed inputs.	A browser or generic fetch tool accepts a large and difficult-to-audit request space.
Interfaces	REST, GraphQL, MCP, and the Skill share one operation catalog.	Each interface may use different schemas, authorization, metering, and error behavior.
Prompt boundary	Marketplace content is returned as untrusted data, not instructions.	Page content can be mixed into an agent context without a reliable trust boundary.
Cost control	A successful operation consumes one published request unit.	Browser duration, tokens, proxy traffic, retries, and extraction steps vary per task.
Freshness	Operation-specific cache policy and fetched time travel with the response.	The agent often cannot distinguish a cached page, fresh page, or replayed result.
Failure states	Validation, authorization, acquisition, and parsing failures remain typed.	The agent may receive partial text, an empty page, or a timeout without a stable error class.

AI agents questions

Which interfaces can an agent use?+

The same operation catalog is available through REST, GraphQL, MCP, and the Glade amazon-data Skill.

Does MCP have a separate authorization model?+

No. MCP resolves through the same organization, API key, operation, metering, and policy boundaries as the other interfaces.

How are prompt injections handled?+

Marketplace and provider text is returned as untrusted response data. It is never treated as a tool description or system instruction.

Can an agent issue arbitrary Amazon requests?+

No. The public tool surface maps to the documented 17-operation catalog and each input is validated before provider work begins.

Can the agent predict call cost?+

Yes. A successful normalized operation consumes one request unit and the response exposes request metadata.

Can I use the same tools outside MCP?+

Yes. REST and GraphQL expose the same canonical operations, which keeps backend jobs and agent calls aligned.

Ask AI about Glade API

Ask ChatGPT Ask Perplexity Ask Gemini Ask Claude Ask Grok

Give your agent a bounded Amazon tool

Use one key and the same typed operation catalog through REST, GraphQL, MCP, or the Glade Skill.

Start free Read the docs

Request

curl -G "https://gladeapi.com/api/amazon/search" \
  -H "API-KEY: glade_live_YOUR_API_KEY" \
  --data-urlencode "searchTerm=mechanical%20keyboard" \
  --data-urlencode "domain=US" \
  --data-urlencode "page=1"

curl -G "https://gladeapi.com/api/amazon/product" \
  -H "API-KEY: glade_live_YOUR_API_KEY" \
  --data-urlencode "asin=B0D1XD1ZV3" \
  --data-urlencode "domain=US"

Both calls use the same key, response envelope, request metadata, and documented failure model.