Connect Claude to Convultra

The Convultra MCP server is in Beta. Connecting Claude takes three clicks in Claude.ai, or one command in Claude Code. Once connected, Claude can read your tracking health, overview, breakdowns, conversions, delivery log and spend, and answer questions about them in plain language.

Claude.ai: add a custom connector

1. Open Settings → Connectors → Add custom connector.
2. Paste https://mcp.convultra.com/mcp.
3. Add it.

Both advanced settings are detected automatically and need no changes: the server always requires authentication, and it advertises support for Anthropic’s hosted client metadata, so Claude registers itself without you configuring anything.

You are then taken to a Convultra consent screen. Sign in, review the plain-English permission lines, and choose all projects or pick individual ones. Approve, and the connector is live.

The permissions Claude asks for are the six read scopes an assistant needs:

Claude asks separately, later, for anything beyond them, such as delivery:write to retry a failed delivery. See API scopes.

Behind the scenes this is OAuth 2.1 with PKCE. Claude discovers Convultra’s authorization server, you sign in, and Claude receives an access token bound to the MCP URL. You never see or paste a token. To disconnect, remove the connector in Claude’s settings; the grant stops working immediately.

Claude Code: one command

claude mcp add --transport http convultra https://mcp.convultra.com/mcp

Claude Code then authenticates. It can use the same OAuth sign-in as Claude.ai, or a static sk_ key created under Settings → Developer, which suits scripted or shared environments. If you use a key, give it only the read scopes you need and, on an agency account, restrict it to the projects the session should see.

First questions to ask

• “Is tracking healthy on acme.example, and when did the last event arrive?”
• “Give me the last 30 days overview compared with the previous period.”
• “Which sources drove revenue last week, and which had spend?”
• “Why did conversions drop last week?” Claude will check tracking health, the overview, a breakdown and the delivery summary, in that order.
• “List failed deliveries to Google Ads in the last 7 days and explain the reasons.”

Each of those becomes several tool calls, so on a Starter plan (60 requests a minute, 5,000 a day) keep an eye on usage. get_me shows what your credential has.

What Claude sees and does not see

Claude gets the same data as the REST API, with the same redaction: no IP addresses, user agents, fingerprints or coordinates, and page URLs reduced to paths. Conversion rows do include click ids and session ids, which is what lets Claude reconcile your figures against Google Ads. Customer-supplied text such as campaign names appears only in structured data, never in the summary Claude reads first. See Privacy and redaction.

If the connector will not add

• Claude shows a “Connect” prompt every time. The OAuth grant was revoked or expired. Remove the connector and add it again.
• A project is missing. You picked individual projects at consent time. Remove and re-add the connector, choosing all projects, or add the project to the grant.
• Claude says it can only see part of a list. Tool results are trimmed to the top 25 rows by default. Ask Claude to narrow the range or filter rather than fetch more.
• plan_required. The API and MCP server are included on every paid plan. This appears only for agency-wide keys on plans without them.

See The MCP server for what to expect once connected, and Connect ChatGPT, Cursor and your own agents for other clients.