How do I connect my own AI tool?

• The same tools from the same package as the portal chat. With the same account, period, model, post-view setting and data state, portal and MCP read the same data; the data keeps growing, so a later request can contain more.
• Without a model choice the tools use your default model and your post-view setting from “Settings”, like the portal. Every answer states under “model” and “includePostView” what was used. Your AI tool does not need to know the default model.
• Models and the reasons why a conversion is not attributed to any touchpoint appear in words in the answers, as in the portal. The language comes from your request’s “Accept-Language” header (German, English, Spanish or French); if your AI tool sends none, it is the language you last signed in to the portal with, and German if that is unknown too.
• The tools themselves appear in that same language, so in the language of your AI tool or of your account: titles and descriptions exist in German and English, with English for Spanish and French.
• A date without a time is a whole day in your account’s time zone, as start and as end; days: 7 means seven whole days including today. Every result states the days and the zone under “period”.
• The breakdown by channel, campaign or ad gives an AI tool at most the 200 largest rows, so the answer fits into the conversation. “total” states how many there are in all, and “truncated” says whether rows were cut; a sum over the rows is then a partial sum. You get all rows from the read API (/v1/data/breakdown?limit=all) or the CSV export on the “Attribution” page.
• Including “help”: your own AI tool can read the articles of this help centre too.
• And “commerce”: new-customer cost, net revenue, contribution margin, repeat purchases and the product report as under “Commerce & products”. As there, a campaign only narrows the product report.
• Read-only throughout: a model should not be able to change data.
• The tenant is bound to the key; separation is additionally enforced in the database.
• A key can be revoked or rotated at any time.

Data protection before you connect

By allowing access or creating a read key, you instruct us to transfer the data of your account to the tool. Within this authorization, the connected tool can request data for your analyses: metrics, ad costs from your connected ad accounts, individual conversions with the order number in plain text and pseudonymous customer journeys. No tool returns names, email addresses, IP addresses or the pages visited. Orders your shop reported without tracking consent are never returned individually, neither by a tool nor by the read API, and neither is their order number; they are included in counts and revenue totals.

You read the notice on this on the consent page before allowing access and before creating a read key. We store which version you saw; after disconnecting, this record is kept for three years.

You can also give the instruction in advance: with an optional checkbox in the step “Conclude the data processing agreement” or under “Privacy”. The button then shows, instead of the full notice, who gave it, when and in which version, with the wording to expand, along with the tool or key, the account and the scope. With “Allow” or “Create key” you confirm the specific connection. The instruction does not apply to other purposes or recipients. If there is a new version, the button shows the full notice again until someone gives it.

If you revoke the advance instruction under “Privacy”, all read keys and all connections to AI tools of this account end immediately, including those someone else created. The next request from a tool is rejected.

Connecting Claude and ChatGPT

You find the address of the MCP server under “Exports & AI connections”. When you enter it in Claude or ChatGPT, the tool sends you to Metrunio: you sign in, choose the account and allow read access. You do not need a key for this.

1. 01Claude in the browser, the desktop app and on your phone: under “Connectors” in the settings, add a custom connector with the address and connect. You can keep the defaults: “Sign in now” as the authentication and “Use Claude’s published identity” as the OAuth client. “Register automatically” works as well; use it if signing in with the default fails.
2. 02ChatGPT: under “Plugins”, use “Add” → “Add custom MCP server” with the address and choose OAuth as the sign-in. In a chat, select it via “+”.
3. 03Claude Code: claude mcp add --transport http metrunio <address>, then run /mcp in Claude Code and sign in.
4. 04On Metrunio’s consent page, check which account is read and where you return afterwards, read the notice on data transfer and choose “Allow”. If Claude signs in with its published identity, the page also shows where the details of the tool come from, that is claude.ai.

The connection then appears under “Exports & AI connections” among the connected AI tools. You disconnect it there; the tool reads nothing from its next request on. It also ends when the person who allowed it may no longer create keys in the account, and it pauses while the plan does not include the MCP server.

Connecting other tools with a key

Tools without OAuth get a read key from “Exports & AI connections”. It goes into the Authorization: Bearer header of every request.

• Claude Code with a key: claude mcp add --transport http metrunio <address> --header "Authorization: Bearer rk_…"
• Claude Desktop with a key: add a server “metrunio” to the configuration file with the command npx mcp-remote <address> --header "Authorization: Bearer rk_…" and restart Claude.