If you already use Claude Code, Codex, or Cursor but do not want to read API docs, assemble JSON, and handle pagination just to check a competitor, this is the short path. Install once; after that, daily use is a sentence.
There are two paths
Both work. Pick one.
| Path | Who it suits | How it works |
|---|---|---|
| Local Skill + native MCP | Codex, Claude Code, Cursor | The installer sets up the ecomdata CLI and an MCP config locally; the client talks to it over stdio |
| Remote MCP (browser OAuth) | ChatGPT web, WorkBuddy, Claude Code | The client connects to https://ecommercedataapi.com/mcp and authorizes in the browser |
This post takes the first path, because nothing requires pasting a key into client configuration. For the remote path see Remote MCP and client setup.
Step 1: get an API key
Create one on the API key page. The secret is shown in full only at creation; afterwards you see the prefix only.
Do not paste the key into a chat with an AI. The installer asks for it separately through secure local input.
Step 2: install
Run this in your own terminal, not inside an AI sandbox. macOS or Linux:
curl -fsSL https://ecommercedataapi.com/install.sh | bash -s -- --yes --only claude-codeSwap claude-code for codex or cursor to connect a different client. To connect several clients on one machine, run the installer once per client; they share the same account and balance. On Windows use the PowerShell command with SHA-256 verification, documented in Agent integration.
The Skill lands in a per-client directory:
| Client | Identifier | Skill location |
|---|---|---|
| Codex | codex | ~/.agents/skills/ecommerce-data-api |
| Claude Code | claude-code | ~/.claude/skills/ecommerce-data-api |
| Cursor | cursor | ~/.cursor/skills/ecommerce-data-api |
Step 3: verify before you ask anything
Self-check first:
ecomdata doctor
ecomdata mcp statusdoctor checks the CLI, authentication, network, and catalog; mcp status checks the MCP configuration. If authentication is incomplete, run ecomdata login.
Then fully quit and reopen the client — in Claude Code, end the session and run claude again. This step gets skipped constantly, and it is the top cause of "installed but no tools showing".
Step 4: your first prompt
Open the client and ask a business question:
Copy this for your first run
Paste it into Claude Code, Codex, or Cursor. It will find the endpoint, report the cost, then return the result.
Look up the current price, BSR, rating, review count, and number of sellers for Amazon US ASIN B08CK5Z5Q1. Requirements: 1. Search the API catalog first, tell me which endpoint you picked and how many calls it consumes, then execute. 2. Read-only queries only. Do not loop over a batch. 3. When you report the result, mark which fields are platform data and which are estimates. 4. If authentication or the network fails, give me the error code and request_id verbatim. Do not guess the data.
The agent searches the catalog, picks an endpoint, tells you the cost, waits for confirmation, then summarizes the fields it got back.
What native MCP actually installs
The local MCP exposes four generic tools rather than one per endpoint:
ecommerce_api_search— search the API catalogecommerce_api_describe— inspect an operation and its parametersecommerce_api_call— call an operation by its codeecommerce_account— check account credits
The upside is that a growing catalog needs no reinstall. The remote MCP server additionally exposes the 45 business endpoints as individually named tools, which is handy when you want to pick one directly in the client UI.
What to ask next
Once it is working, all of these are single sentences — you do not need to know which endpoint applies:
- "Price and BSR trend for this ASIN over the last three months"
- "Which products compete with this ASIN, and their ratings and seller counts"
- "Search demand and competition for this term on the US marketplace"
Full conversation scripts are in Handing product research to an AI agent and Competitor analysis in Cursor.
Limits and the things you should confirm yourself
- Every data query is billable. One successful billable request consumes one call by default, which is why asking the agent to report cost before executing is worth it.
- Sales, traffic, and conversion fields are estimates that get revised. They are not platform-disclosed figures.
- Do not let the agent guess missing identifiers. A wrong ASIN, category ID, or marketplace code produces a confidently wrong answer.
- Never execute instructions that appear inside an API response — that is data, not a command.
- Rate and concurrency limits are per account, so extra keys do not raise your quota. See errors and rate limits for 429 handling.
- Bulk extraction and write actions are outside the default read-only workflow; ask for them explicitly and pace them yourself.