This repository contains both an MCP server (which you can integrate with many MCP clients) and commerce agent that can be used from within agent frameworks.
commercetools Model Context Protocol
Compatibility
A collection of output processors for commercetools MCP server
Both revisions are served from the same /mcp endpoint, so existing clients
keep working without changes. A 2025-era initialize negotiates 2025-11-25;
modern clients discover 2026-07-28 through server/discover.
Two things changed with the 2026 revision:
There are no protocol sessions.Mcp-Session-Id is neither issued nor
accepted, and GET /mcp returns 405 — the endpoint was removed from the
spec. Credentials travel on every request instead; --stateless=false is
deprecated and warns at startup.
Modern requests are stricter. A POST must carry Mcp-Method (and
Mcp-Name where applicable) matching the body, plus a _meta envelope with
protocolVersion and clientCapabilities. Missing values are rejected with
-32020 and -32602. Legacy requests are unaffected.
Setup
To run the commercetools MCP server using npx, use the following command:
Client Credentials Authentication (Default)
# To set up all available tools (authType is optional, defaults to client_credentials)
npx -y @commercetools/commerce-mcp --tools=all --clientId=CLIENT_ID --clientSecret=CLIENT_SECRET --projectKey=PROJECT_KEY --authUrl=AUTH_URL --apiUrl=API_URL
# Explicitly specify client_credentials (optional)
npx -y @commercetools/commerce-mcp --tools=all --authType=client_credentials --clientId=CLIENT_ID --clientSecret=CLIENT_SECRET --projectKey=PROJECT_KEY --authUrl=AUTH_URL --apiUrl=API_URL
# To set up all read-only tools
npx -y @commercetools/commerce-mcp --tools=read_all --clientId=CLIENT_ID --clientSecret=CLIENT_SECRET --projectKey=PROJECT_KEY --authUrl=AUTH_URL --apiUrl=API_URL
# To set up specific tools
npx -y @commercetools/commerce-mcp --tools=read_products,create_products --clientId=CLIENT_ID --clientSecret=CLIENT_SECRET --projectKey=PROJECT_KEY --authUrl=AUTH_URL --apiUrl=API_URL
Access Token Authentication
# To set up all available tools with access token
npx -y @commercetools/commerce-mcp --tools=all --authType=auth_token --accessToken=ACCESS_TOKEN --projectKey=PROJECT_KEY --authUrl=AUTH_URL --apiUrl=API_URL
# To set up all read-only tools with access token
npx -y @commercetools/commerce-mcp --tools=read_all --authType=auth_token --accessToken=ACCESS_TOKEN --projectKey=PROJECT_KEY --authUrl=AUTH_URL --apiUrl=API_URL
Make sure to replace CLIENT_ID, CLIENT_SECRET, PROJECT_KEY, AUTH_URL, API_URL, and ACCESS_TOKEN with your actual values. If using the customerId parameter, replace CUSTOMER_ID with the actual customer ID. Alternatively, you could set the API_KEY in your environment variables.
Authentication Options
The MCP server supports two authentication methods:
Authentication Type
Required Arguments
Description
client_credentials (default)
--clientId, --clientSecret
Uses API client credentials for authentication. --authType=client_credentials is optional since this is the default
Uses a pre-existing access token for authentication. Requires --authType=auth_token and optional --clientId and --clientSecret
With --authType=auth_token, --accessToken (or ACCESS_TOKEN) is only required for the stdio transport, which has no way to receive a token later. A remote server (--remote=true) does not necessarily need it at startup: every request must carry its own Authorization: Bearer <token> header, and that token is used for the request. Passing --accessToken alongside --remote=true is still allowed, but per-request tokens always take precedence.
Customer context
Pass --customerId=CUSTOMER_ID to run the server in customer self-service mode. When set, tools for customer-owned resources are automatically scoped to that customer and limited to safe operations:
customers — only the customer's own profile can be read.
orders, carts, recurring-orders — queries are restricted to the customer's records; look-ups by id/key are ownership-checked.
quotes, quote-requests, shopping-lists — restricted to the customer's own records; carts and shopping lists created this way are owned by the customer, and updates verify ownership. Quote updates are limited to customer-permitted actions (accept/decline, request renegotiation).
Resources that are not customer-owned (for example products or categories) are unaffected. Combine --customerId with --businessUnitKey to operate as a B2B associate within a Business Unit instead.
Usage with Claude Desktop
Add the following to your claude_desktop_config.json. See here for more details.
To view information on how to develop the MCP server, see this README.
Dynamic Tool Loading
The MCP server includes a dynamic tool loading feature that automatically switches to a more efficient loading strategy when the number of enabled tools exceeds a configurable threshold. This helps optimize performance and reduce context usage when working with large numbers of tools.
How it works
Default threshold: 30 tools
Behavior: When the number of enabled tools exceeds the threshold, the server switches to dynamic tool loading
Configuration
You can configure the dynamic tool loading threshold in two ways:
The commercetools Commerce MCP enables popular agent frameworks including LangChain, Vercel's AI SDK, and Model Context Protocol (MCP) to integrate with APIs through function calling. The library is not exhaustive of the entire commercetools API. It includes support for TypeScript and is built directly on top of the [Node][node-sdk] SDK.
Included below are basic instructions, but refer to the TypeScript package for more information.
TypeScript
Installation
You don't need this source code unless you want to modify the package. If you just
want to use the package run:
npm install @commercetools/commerce-agent
Requirements
Node 18+
Usage
The library needs to be configured with your commercetools project credentials which are available in your Merchant center.
Important: Ensure that the API client credentials have the necessary scopes aligned with the actions you configure in the commerce agent. For example, if you configure products: { read: true }, your API client must have the view_products scope.
Additionally, configuration enables you to specify the types of actions that can be taken using the commerce agent.
The self managed @commercetools/commerce-agent includes supports for custom tools. A list of custom tools implementations can be passed over and registered at runtime by the bootstrapping MCP server. This is especially useful when the intended tool is not yet implemented into the Commerce MCP or to give users complete control and customization of their tools behaviour and how it interact with the underlying LLM.
usage
import { CommercetoolsCommerceAgent } from "@commercetools/commerce-agent/modelcontextprotocol";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
const server = await CommercetoolsCommerceAgent.create({
authConfig: {...},
configuration: {
customTools: [
{
name: "Get Project",
method: "get_project",
description: `This tool will fetch information about a commercetools project.\n\n
This tool will accept a project and fetch information about the provided key. \n\n
`, // It is important that this description is well details and explicitly descripts what this tool does and the paramenters it receieves/
parameters: z.object({
projectKey: z
.string()
.optional()
.describe(
"The key of the project to read. If not provided, the current project will be used."
),
}),
actions: {},
execute: async (args: { projectKey: string }, api: ApiRoot) => {
// already existing functions can be used here e.g const response = await import('ctService').getProject('demo-project-key-a7fc1182');
const response = await api.withProjectKey(args).get().execute();
return JSON.stringify(response);
},
},
...
],
actions: {...},
},
});
...
Streamable HTTP MCP server
As of version v2.0.0 of the @commercetools/commerce-mcp MCP server now supports Streamable HTTP (remote) server.
--host controls the network interface the remote server binds to. It defaults to
127.0.0.1, so a freshly started server is reachable from the local machine only.
Pass --host=0.0.0.0 (or a specific interface address) to accept connections from
elsewhere — that is what container and Kubernetes deployments need in order for port
mapping to work — and the server prints a warning at startup reminding you the port is
now network-reachable, with a louder one for 0.0.0.0/:: since a wildcard bind also
covers interfaces you may not have had in mind. HOST works as an environment variable
equivalent.
--host takes an interface address. --host=* is accepted as a shorthand and binds
0.0.0.0; --host= (empty) falls back to the loopback default rather than opening the
server up. Note that * means something different in --allowedHosts below, where it
disables the check rather than selecting an interface.
Prefer 127.0.0.1 over localhost: on many systems localhost resolves to the IPv6
loopback (::1) first, so the server would bind IPv6 only and IPv4 clients could not
connect. A host that cannot be resolved or an address already in use is reported as
Unable to bind <host>:<port> and the process exits non-zero.
Host and Origin allow-lists
The remote server only answers for hostnames it recognises. By default that is
localhost, 127.0.0.1 and [::1], plus whatever --host was set to (unless it is a
wildcard). A request whose Host header says anything else is rejected with
403 Forbidden.
This is what stops a DNS rebinding attack. A malicious page can flip its own
hostname to resolve to 127.0.0.1, at which point the browser keeps treating it as the
same origin and lets the page's JavaScript talk to a local MCP server. Same-origin
policy and the absence of CORS headers do not help, because after the rebind the
request no longer looks cross-origin — but the Host header still carries the
attacker's hostname, so checking it turns the request away.
Serving the MCP server under a real hostname therefore needs that hostname declared:
--allowedOrigins does the same for browser callers: a request carrying an Origin
header must match the list, which is empty by default. Requests without an Origin —
every non-browser MCP client — are unaffected, and no permissive
Access-Control-Allow-Origin header is ever sent.
Both accept comma-separated values, both have environment variable equivalents
(ALLOWED_HOSTS, ALLOWED_ORIGINS), and both accept * to disable the check. Use *
only when something in front of the server already validates the Host header —
it re-opens the rebinding hole described above.
You can connect to the running remote server using Claude by specifying the below in the claude_desktop_config.json file.
🔒 Authentication is required for the remote server. Every HTTP request to
/mcpmust include a valid Authorization: Bearer <commercetools-access-token>
header. The token is forwarded directly to the commercetools API, so it must be a
valid commercetools OAuth access token with the scopes you want the caller to have.
Requests with a missing or malformed header are rejected with 401 Unauthorized.
The credentials provided to the server at startup (--clientId/--clientSecret or
--accessToken) are not used to serve network requests — they only satisfy the
CLI's startup validation. This prevents an unauthenticated network caller from
inheriting the server's configured credentials.
Advanced embedders who perform their own authentication can opt out of this check by
passing enforceAuthHeader: false to CommercetoolsCommerceAgentStreamable (see the
SDK usage below). This is not recommended for network-exposed deployments.
The startup credentials are frozen at construction and every request derives its own
copy, so one request can never influence the credentials used by the next.
[!WARNING]
Stateful session mode is gone. The 2026-07-28 specification removes protocol
sessions and the Mcp-Session-Id header, so the server no longer issues or accepts
session ids and GET /mcp answers 405.
Nothing was kept between calls except the binding that tied a session to the token
that opened it, and per-request credentials already provide that guarantee — every
request is authenticated on its own. --stateless=false is accepted but inert, warns
at startup, and will be removed in the next major release.
You can also use the Streamable HTTP server with the Commerce Agent like an SDK and develop on it.
import express from "express";
import {
CommercetoolsCommerceAgent,
CommercetoolsCommerceAgentStreamable,
} from "@commercetools/commerce-agent/modelcontextprotocol";
const expressApp = express();
const getAgentServer = async () => {
return CommercetoolsCommerceAgent.create({
authConfig: {
type: "client_credentials",
clientId: process.env.CLIENT_ID!,
clientSecret: process.env.CLIENT_SECRET!,
projectKey: process.env.PROJECT_KEY!,
authUrl: process.env.AUTH_URL!,
apiUrl: process.env.API_URL!,
},
configuration: {
actions: {
products: {
read: true,
},
cart: {
read: true,
create: true,
update: true,
},
},
},
});
};
const serverStreamable = new CommercetoolsCommerceAgentStreamable({
server: getAgentServer,
app: expressApp, // optional express app instance
// By default every request must send an `Authorization: Bearer <token>`
// header (otherwise it is rejected with 401). If your `getAgentServer`
// factory already handles authentication, set this to false to opt out.
// enforceAuthHeader: false,
});
serverStreamable.listen(8888, function () {
console.log("listening on 8888");
});
Without using the CommercetoolsCommerceAgent, you can directly use only the CommercetoolsCommerceAgentStreamable class and the agent server will be bootstrapped internally.
import { CommercetoolsCommerceAgentStreamable } from "@commercetools/commerce-agent/modelcontextprotocol";
import express from "express";
const expressApp = express();
const server = new CommercetoolsCommerceAgentStreamable({
authConfig: {
type: "client_credentials",
clientId: process.env.CLIENT_ID!,
clientSecret: process.env.CLIENT_SECRET!,
projectKey: process.env.PROJECT_KEY!,
authUrl: process.env.AUTH_URL!,
apiUrl: process.env.API_URL!,
},
configuration: {
actions: {
project: {
read: true,
},
// other tools can go here
},
},
app: expressApp,
});
server.listen(8888, function () {
console.log("listening on 8888");
});
