Connect Claude, ChatGPT and Cursor

AxiTrace runs a remote MCP server. Connect it to Claude, ChatGPT, Cursor or any other MCP client and ask in plain language about your store: revenue and ROAS, where orders come from, which campaigns sell, which source or campaign brought the most add-to-carts that did not end in an order, whether every ad platform connection is healthy. The assistant reads the same numbers as your AxiTrace dashboard. If you allow it, it can also set up custom events for you; everything else is read-only.

On this page

• Before you start
• How connecting works
• Claude.ai and Claude Desktop
• Claude Code
• ChatGPT (official plugin, desktop app or chatgpt.com)
• Cursor
• Other MCP clients
• Personal access tokens
• The MCP tools
• Setting up custom events
• What data is and is not exposed
• Plans
• Rate limits
• Security e-mails
• Revoking access
• Troubleshooting

Before you start

• A workspace on a paid plan (Pro, Business or Custom; existing legacy Starter also retains access). The Free plan does not include API access. The two fraud tools also need the Business or Custom plan.
• Your own AxiTrace login with access to that workspace. The assistant can read only workspaces you can see yourself, and it loses access to a workspace when you do.
• An assistant that supports remote MCP servers over HTTP, with OAuth sign-in or a custom Authorization header.

How connecting works

1. You add https://axitrace.com/api/mcp to your assistant as a custom connector (or MCP server).
2. The assistant finds the AxiTrace sign-in details on its own and opens an AxiTrace page in your browser. Sign in if you are not signed in yet.
3. The consent page shows the name of the application, what it will be able to read, and the workspaces it may read. All eligible workspaces are selected; untick the ones you do not want to share. At least one must stay selected. Only workspaces on a paid plan are offered, and demo workspaces are marked Demo.
4. Click Approve. You return to the assistant, which can now call the tools.

A few things to know about a connection:

• Workspaces you create later are not included. Connect the application again to add them.
• A connection lasts up to 180 days from your approval. The assistant renews its short-lived access token (one hour) in the background, so you do not notice it. After 180 days, connect again.
• You receive a security e-mail for every new connection.
• You can disconnect at any time: see Revoking access.

Claude signs you in with OAuth, so you need no token. AxiTrace is listed in the Claude Connectors Directory, so the quickest way is to connect it from there.

From the Claude Connectors Directory (recommended)

1. Open AxiTrace in the Claude Connectors Directory while signed in to claude.ai.
2. Click Connect.
3. The AxiTrace page Connect Claude to AxiTrace opens (sign in first if asked). Untick the workspaces you do not want to share, tick any write permission you want to allow and click Approve.
4. Open a new chat and ask, for example, "List my AxiTrace workspaces". The first time Claude uses each AxiTrace tool it asks for permission: click Allow once or Always allow.

Once connected, the listing shows Connected and AxiTrace appears under Customize, Connectors on claude.ai.

As a custom connector

You can also add the same server by hand. The steps below were checked on claude.ai on 24 September 2026.

1. On claude.ai, click Customize in the left sidebar and open the Connectors tab.
2. Click + Add (top right) and choose Add custom connector.
3. Step 1 of the dialog: Name AxiTrace, MCP server URL https://axitrace.com/api/mcp. Click Continue.
4. Step 2: Claude reads the AxiTrace sign-in details on its own and marks two options as Detected. Keep them exactly as they are:
   • Authentication: Sign in now;
   • OAuth client: Use Claude's published identity;
   • Request headers: add nothing. Do not paste a token or an Authorization header here: Claude gets its own token when you sign in, and a header of yours would get in its way.
   Click Add.
5. The connector page says You're not connected to AxiTrace yet. Click Connect.
6. The AxiTrace page Connect Claude to AxiTrace opens (sign in first if asked). It lists what Claude will be able to read and your workspaces, all ticked. Untick the ones you do not want to share and click Approve.
7. You are back on claude.ai with the message Connected to AxiTrace.
8. Open a new chat and ask, for example, "List my AxiTrace workspaces" or "AxiTrace overview for the last 7 days". The first time Claude uses each AxiTrace tool it asks Claude wants to use ... from AxiTrace: click Allow once, or Always allow to stop being asked for that tool.

A connector added on claude.ai is available in Claude Desktop and the Claude mobile apps under the same Claude account. If your organization manages connectors centrally, an owner may have to add the connector before members can connect it.

Option A: sign in with OAuth (recommended for your own machine).

claude mcp add --transport http axitrace https://axitrace.com/api/mcp

Then run /mcp inside Claude Code, pick axitrace and choose to authenticate. Your browser opens the AxiTrace consent page. The page notes that the authorization code goes to an application running on your own computer: that is Claude Code, which is expected.

Option B: a personal access token (for shared setups, CI or when a browser is not available). Create a token first (see Personal access tokens) and keep it in an environment variable:

export AXITRACE_TOKEN="paste the token here"
claude mcp add --transport http axitrace https://axitrace.com/api/mcp --header "Authorization: Bearer $AXITRACE_TOKEN"

Your shell expands $AXITRACE_TOKEN when you run the command, so the token ends up in the Claude Code configuration. For a project shared through git, use a .mcp.json file that references the variable instead, and never commit the token itself:

{
  "mcpServers": {
    "axitrace": {
      "type": "http",
      "url": "https://axitrace.com/api/mcp",
      "headers": {
        "Authorization": "Bearer ${AXITRACE_TOKEN}"
      }
    }
  }
}

AxiTrace is an official ChatGPT plugin, approved by OpenAI, so the quickest way is to install it from the ChatGPT plugin directory.

From the ChatGPT plugin directory (recommended)

1. Open AxiTrace in the ChatGPT plugin directory while signed in to chatgpt.com.
2. Click Install plugin.
3. The AxiTrace page Connect ChatGPT to AxiTrace opens (sign in first if asked). Untick the workspaces you do not want to share, tick any write permission you want to allow and click Approve.
4. Open a new chat and ask, for example, "Use AxiTrace: which traffic sources brought the most revenue last month?"

Adding the server by hand

You can also add the same server yourself. ChatGPT has two different places for that, and they do not look alike. Pick the one that matches where you use ChatGPT. Either way you need a Plus, Pro, Business, Enterprise or Education account.

Option A: the ChatGPT desktop app

The desktop app connects AxiTrace with a personal access token. Create the token first and keep it at hand: AxiTrace shows it only once.

1. Click Plugins in the left sidebar.
2. Click Add (the dark button, top right, next to Browse catalog) and choose the MCP server option. The page Connect to a custom MCP opens.
3. Fill in the form exactly as in the table below. Leave every field that the table marks as empty untouched.
4. Click Save. AxiTrace now appears under the MCP servers tab of the Plugins page.
5. Open a new chat (an existing one does not see the new server) and ask "List my AxiTrace workspaces". The answer lists your workspaces with their plan, currency and timezone.

To change the URL or the token later, open AxiTrace under MCP servers: the page Update MCP AxiTrace shows the same fields. Never share a screenshot of it with the token visible.

Option B: ChatGPT in a browser (chatgpt.com)

On the web, ChatGPT connects AxiTrace in developer mode and signs you in with OAuth, so no token is needed. The steps below were checked on chatgpt.com on 24 September 2026. Button names are given as in the English interface; in another language they sit in the same places.

1. Turn on developer mode (once). On chatgpt.com, click your name in the bottom left corner, open Settings and choose Security and login. In the Developer mode section, turn on the Developer mode switch (ChatGPT marks it as elevated risk; every AxiTrace tool is read-only except custom event and survey authoring tools, which work only if you allow them). Close the settings and reload the page (Cmd+Shift+R on a Mac, Ctrl+Shift+R on Windows).
2. Click Plugins in the left sidebar, then the plus button next to the Search plugins box. From its menu choose Create app.
3. The New plugin dialog opens with a drop zone for plugin files. Ignore it and click Create MCP app in its bottom left corner.
4. Fill in the form:

   Field	What to enter
   Icon (optional)	Leave empty.
   Name	AxiTrace
   Description (optional)	Leave empty.
   Connection	Keep Server URL (not Tunnel) and enter https://axitrace.com/api/mcp.
   Authentication	OAuth (preselected).
   Advanced OAuth settings	Leave as they are: ChatGPT reads them from AxiTrace.
   I understand and want to continue	Tick it (it is ChatGPT's general warning about custom servers).

   Click Create.
5. ChatGPT shows Add AxiTrace to ChatGPT. Click Sign in with AxiTrace. A new browser tab opens the AxiTrace page Connect ChatGPT to AxiTrace: choose the workspaces and click Approve. The tab closes and ChatGPT confirms AxiTrace account connected.
6. Load the tools (easy to miss). Right after connecting, the app's settings page says there are no tools yet. Scroll to Information and click Refresh; ChatGPT confirms Tools refreshed and lists the AxiTrace tools. Without this step ChatGPT answers without AxiTrace.
7. Open a new chat and name the app in your question, for example "Use the AxiTrace app: list my workspaces, then show the 3 top campaigns by revenue in the last 30 days".

To find the app again later: Plugins → the Personal tab → Created by me → AxiTrace; the ... menu → Manage opens its settings (connected account, tools, Refresh). If the plus button does nothing, reload the page: ChatGPT shows it only once developer mode is on.

No Developer mode switch under Security and login? On a Business, Enterprise or Education workspace an admin has to allow it first (Workspace settings → Permissions & roles → Developer mode). OpenAI renames these menus from time to time; its developer mode guide has the current path.

The server marks report tools as read-only and authoring tools as writes. Grant custom_events:write or surveys:write explicitly to enable those operations. Your assistant controls how it presents tool confirmations.

Add AxiTrace to ~/.cursor/mcp.json (all projects) or .cursor/mcp.json (one project) with a personal access token read from an environment variable:

{
  "mcpServers": {
    "axitrace": {
      "url": "https://axitrace.com/api/mcp",
      "headers": {
        "Authorization": "Bearer ${env:AXITRACE_TOKEN}"
      }
    }
  }
}

Set AXITRACE_TOKEN in the environment Cursor starts from, then check that the axitrace server and its tools are listed in Cursor's MCP settings. Pasting the token directly into mcp.json also works, but then the file itself is a secret and must never be committed.

Other MCP clients

Any client that speaks the MCP Streamable HTTP transport can connect:

• Endpoint: POST https://axitrace.com/api/mcp (Streamable HTTP; GET is not offered, the server does not stream on its own).
• OAuth 2.1: the first unauthenticated request answers 401 with a WWW-Authenticate header pointing to https://axitrace.com/.well-known/oauth-protected-resource/api/mcp. Authorization code with PKCE (S256), client registration through a Client ID Metadata Document or dynamic client registration, and resource=https://axitrace.com/api/mcp.
• Or a personal access token as Authorization: Bearer <token>.

The OAuth details are the same as for the REST API: see REST API v1, Authentication.

Personal access tokens

A personal access token is a credential you create yourself, for Claude Code, Cursor, the ChatGPT desktop app, scripts and anything else that cannot sign in with OAuth. It works for both the MCP server and the REST API.

1. Open the account menu and choose AI assistants and API (/admin/account/api-access).
2. Under Create a personal access token, give it a label you will recognise later (for example Cursor on my laptop), keep the scopes and workspaces it needs, and pick how long it lives: 30, 90, 180 or 365 days (90 by default). Click Create token.
3. Copy the token right away. It is shown only once. It is a long string (about 900 characters).

The MCP tools

Reports read data. Custom event tools require the opt-in custom_events:write scope (see Setting up custom events); survey authoring tools require surveys:write; settings and Product Labels changes require settings:write Fraud Detection actions fraud:write and excluding orders from the statistics transactions:write (see Settings API and MCP and Fraud Actions API and MCP). No tool changes your tracked data, orders, connections or campaigns (set_order_exclusion only takes orders out of the statistics and keeps them): the one change AxiTrace can make in an ad account is the Google Ads IP exclusion of an address you block while Fraud Detection Auto Exclusion is on. Call list_workspaces first; the other tools take a workspace_id from it, which may be left out when the connection covers exactly one workspace.

Date ranges. Reports take a preset (last_7_days, last_14_days, last_30_days, last_90_days, last_180_days; the default is last_30_days) or a custom from and to date at most 180 days apart. The fraud tools accept only last_7_days or last_30_days. A preset works like the same preset on the dashboard: last_7_days runs from the same weekday one week ago up to and including today, so today's figures are still partial. Dates follow the workspace's reporting timezone, and money is always an amount with its currency.

Profit and POAS. The profit:read scope is never granted by default. On the consent page it is an extra box you must tick, and on the personal access token form it starts unticked; a connection or token made before it existed does not have it, so reconnect or create a new token to add it. Even with the scope, the profit fields appear only when profit tracking is on for the workspace and the workspace owner has not hidden profit from your account; otherwise they are left out of the answer entirely. Product and line costs are never sent to an assistant.

Add-to-cart figures. If carts and carts_without_order cannot be read for a request, the traffic source or campaign report still answers: both fields are empty (null) and carts_unavailable_reason says why: the workspace is busy, the read failed, the period has too many ad tracking parameter combinations (pick a shorter period), or, for campaigns, the figures of the period are still being prepared in the background (ask again later) or the period starts more than 180 days ago (campaign add-to-carts cover the last 180 days; pick a later period).

When a tool cannot answer (a missing scope, a Free workspace, a range that is too long), the assistant receives one sentence that says what to do next, for example to reconnect or to pick another range.

Setting up custom events

An assistant can create custom events for you, for example "Track clicks on the size guide button as a custom event and send it to GA4". It uses the same rules as the Custom Events page of the dashboard and shows up there like any other event.

• You have to allow it. The custom_events:write scope is an extra box on the consent page that starts unticked, and on the personal access token form it starts unticked too. A connection made without it, or before it existed, stays read-only: reconnect and tick it to add it.
• Your role must allow editing the workspace settings (the owner, or a team member with edit rights). Demo workspaces cannot be changed.
• Your plan sets the limit. Custom events are included in the Pro, Business and Custom plans, each with a limit of active custom events.
• Only connected platforms can be destinations. Connect a platform in AxiTrace first.
• Deleting stays in the dashboard. An assistant can switch an event off, but not delete it, because a deleted key can never be used again.

Assistants ask you before they call a tool that changes something. Check the key, the CSS selector and the destinations before you confirm, then try the trigger on your store with the selector tester.

Surveys from your AI assistant

With the opt-in surveys:read scope, use list_surveys, get_survey, list_survey_templates, get_survey_template and get_survey_results. Results include visitor-written text, so review what you share. With surveys:write, create/edit drafts and templates, add/edit/remove questions, set every builder field, preview and activate/pause/end surveys, configure rewards and import coupon pools.

Start with: Use AxiTrace to create a draft survey for all product pages, asking what information would help visitors decide. Show me the questions and a preview before activating it. The Survey API and MCP reference lists all 23 survey tools, fields and examples. After the first response the content is locked; duplicate the survey to revise it. Both survey scopes start unticked. Reconnect, grant them and refresh tools to enable an existing assistant.

Leads from your AI assistant

With the opt-in leads:read scope, use list_leads (every filter of the Leads page), get_lead (one lead with its products, attribution, conversion and notes), get_lead_kpis, list_lead_statuses and get_lead_sources. Contact details, the inquiry message and the text of the team's notes need the separate opt-in lead_contacts:read scope and a role that may see personal data. With leads:write, the assistant can add notes, change a lead's status, exclude or restore a lead, create an order from a lead and link an existing order.

Start with: Which leads from Facebook this week have not converted yet? or Draft an e-mail to this lead based on their inquiry and the products they asked about. The Leads API and MCP reference lists all 10 lead tools, filters and examples. The three lead scopes start unticked. Reconnect, grant them and refresh tools to enable an existing assistant.

What data is and is not exposed

An assistant can read:

• Only the workspaces you selected when connecting, and only with the scopes shown on the consent page.
• Report figures: KPIs, traffic sources, campaigns, orders with their value and attributed source, integration health, custom event definitions and, on Business and Custom plans, fraud results.
• With the profit:read scope, profit, POAS (Profit on Ad Spend) and cost coverage per report, source, campaign and order, if profit tracking is on and the owner has not hidden it from you.
• With the journeys:read scope, the visits and ad clicks behind each order: times, channels, UTM parameters, click ids with their values, landing URLs and referrers.

An assistant never receives:

• Structured buyer contact details are not exposed. With surveys:read, visitor-written answers are shared as entered and may contain personal information. The one exception you choose yourself: with lead_contacts:read, the contact details your leads entered (e-mail, phone, name, inquiry message) and the text of your team's lead notes, for a role that may see personal data.
• Visitor IP addresses, visitor or session identifiers, except those of a lead under lead_contacts:read. Fraud results show addresses masked to their network (IPv4 /24, IPv6 /48).
• Product costs, line costs or the cost settings of profit tracking.
• Stored credentials: no ad platform tokens, API keys or endpoint secrets. Explicit reward-secret rotation returns the newly generated secret once to its authorized caller.
• Any way to change your tracked data, existing orders, connections or campaigns. Write tools manage custom events, surveys, leads, workspace settings, Product Labels rules and Fraud Detection only with the scopes you tick yourself; create_lead_order records a new order for a lead and forwards it to ad platforms only if you ask for it, and a blocked IP address reaches your Google Ads account as an IP exclusion only while Fraud Detection Auto Exclusion is on.

Every text value that comes from your store or ad accounts (source, campaign and event names, click ids, landing URLs, workspace names, order ids) can contain any text; AxiTrace tells the assistant to treat it as data, never as instructions. E-mail addresses found in these report texts, for example in a landing URL, are removed before the answer leaves AxiTrace. Survey content and visitor-written answers are returned as entered.

When you connect an assistant, the reports it fetches are handled by that AI provider under your own agreement with them. AxiTrace does not send your data to any AI provider on its own initiative.

Plans

• Pro, Business and Custom (plus existing legacy Starter): report tools other than the fraud tools. Custom event and survey authoring and survey results require Pro, Business or Custom.
• Business and Custom: also the fraud tools get_fraud_summary, list_fraud_verdicts, get_fraud_timeline, list_refund_claims and get_refund_claim.
• Free: no API access. A Free workspace still appears in list_workspaces of an existing connection with plan_required set, and its reports are refused until the plan is upgraded.

Rate limits

When a limit is reached, the assistant is told how many seconds to wait. Normal conversations stay far below these numbers. See REST API v1, Rate limits for the headers.

Security e-mails

AxiTrace e-mails you every time:

• a new application is connected to your account, and
• a new personal access token is created.

The e-mail names the application or token, the workspaces it can read, when it happened and from which IP address and browser, and links to the AI assistants and API page. If it was not you, revoke the access there right away and change your password.

Revoking access

Open AI assistants and API in the account menu (/admin/account/api-access). It lists your connected applications and personal access tokens, with the workspaces and scopes of each, when it was last used and its recent calls. Click Revoke next to one: it stops working immediately, and the application gets an authentication error on its next call.

Removing the connector inside Claude, ChatGPT or Cursor does not necessarily revoke it on the AxiTrace side, so revoke it in AxiTrace too.

Troubleshooting

Still stuck? Write to info@axitrace.com with the name of your assistant and the time of the failing call. Never include your token.