Use Joinsight in your AI assistant

Joinsight runs a remote MCP server, so you can ask Claude, ChatGPT, Gemini, Microsoft Copilot, Mistral Le Chat, Perplexity, Cursor, VS Code, Windsurf or any other client that supports remote MCP servers about your revenue, and get the same numbers the dashboard shows. The server is read-only. It serves pre-computed metrics from the same functions behind the dashboard screens. The model's job is to pick the right tool and explain what comes back, not to compute MRR its own way.

This page covers how to connect, what happens when you approve a connection, the tools and what they return, and how to revoke access. For the reasoning behind a metrics server like this, see Ask Claude about your revenue. The product overview is on the MCP server page.

The server URL

The server URL is your Joinsight address followed by /mcp:

https://joinsight.app/mcp

You can also copy it from Settings → AI assistants, under "Connect an AI assistant". The URL is the same for every workspace. You choose the workspace when you approve the connection, not in the URL.

Add it to your assistant

Menu names change between client versions, but the flow is the same everywhere. You add the URL, a browser window opens on Joinsight, you sign in and approve. There is no API key to paste.

Claude

In Claude, open Settings, then Connectors, and choose Add custom connector. Paste the server URL, then sign in to Joinsight and pick the workspace. In a new conversation, check that the connector is turned on before you ask.

In Claude Code, add it from the terminal:

claude mcp add --transport http joinsight https://joinsight.app/mcp

ChatGPT

ChatGPT adds custom MCP servers through developer mode. Open Settings, then Apps, and turn on developer mode under Advanced. Create an app with the server URL and choose OAuth as the authentication. ChatGPT then sends you to Joinsight to approve it.

Cursor

Add the server to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "joinsight": {
      "url": "https://joinsight.app/mcp"
    }
  }
}

Then select Connect next to joinsight in Cursor's MCP settings and approve it in the browser.

Gemini

In the Gemini web app, open Settings, then Connected apps, and add a custom app with the server URL. Custom apps are only offered on some accounts and regions, and only in English for now; once added, the app also shows in the Gemini mobile app. On Gemini Enterprise, an admin adds Joinsight as a custom MCP server for the team.

Microsoft Copilot

In Copilot Studio, open your agent, add a tool and choose Model Context Protocol. Paste the server URL and pick OAuth with dynamic discovery: Copilot Studio finds Joinsight's sign-in from the server and registers itself, so there is no client ID or secret to enter.

Mistral Le Chat

Open Intelligence, then Connectors, choose Add connector and the Custom MCP connector tab. Paste the server URL. Le Chat detects OAuth and opens Joinsight to approve the connection. On team plans, adding connectors may be limited to admins.

Perplexity

On a paid plan, open Settings, then Connectors, and choose Custom connector. Paste the server URL and sign in. Enable the connector in each thread where you want to use it.

VS Code

Run the command MCP: Add Server, choose HTTP and paste the server URL. VS Code opens Joinsight to approve the connection, and GitHub Copilot's agent mode can then call the tools. To share the setup with a project, add it to .vscode/mcp.json:

{
  "servers": {
    "joinsight": {
      "type": "http",
      "url": "https://joinsight.app/mcp"
    }
  }
}

Windsurf

Add the server to Windsurf's MCP config, ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "joinsight": {
      "serverUrl": "https://joinsight.app/mcp"
    }
  }
}

Other clients

The server uses the Streamable HTTP transport with OAuth, and it answers each request with plain JSON. It has no sessions and no server-sent event stream, so a client that only supports the older SSE transport can't connect. Any client that discovers OAuth from the server's 401 response, as the MCP specification describes, should work.

Approving the connection

When the client sends you to Joinsight, you sign in if you aren't already, and see a consent screen titled "Connect" followed by the app's name. It shows:

• Sends you back to: the host the app will receive the authorization on.
• Identified by: for apps that publish a client metadata document, the host of that document. For apps that registered themselves on the fly, it says "Self-registered. Continue only if you started this from" the app's name. A self-registered app chooses its own name, so only approve it if you just started the connection yourself.
• It can: read MRR, movements and customers. Read page, keyword and App Store revenue. Read alerts and data health. See customer emails only masked.
• It can't: change any data or settings. See your API keys or tokens. See full customer emails. Touch billing or members.
• Workspace: a list of the workspaces you belong to. The connection reads exactly one.

Choose "Allow read access" to approve or "Deny" to refuse. The token the app receives is bound to that workspace, to you and to that app. It can't see any other workspace, even one you own. To ask about a second workspace, connect again and choose it there.

The tools

There are 14 tools. Every one is marked read-only to the client. Inputs with a default can be left out. Period inputs take a key (7d, 28d, 30d, 90d, 12m, this_month, last_month) or explicit dates as {start, end} in YYYY-MM-DD, inclusive, in UTC. An end date in the future is moved back to today, and a period can span at most 3 years.

get_social_posts is only listed in workspaces where Instagram insights are turned on, so a workspace without them sees 13 tools. The demo workspace has them on.

Measured and modeled

Keyword revenue and keyword opportunities are modeled. Google never tells anyone which query a specific visitor typed, so revenue is split across queries by their share of a page's clicks. Those responses have modeled: true, and the server tells the model to call the figures estimates. Everything else is measured. How attribution works explains the difference.

Why explain_change matters

explain_change is the tool that answers "why". It runs the same diff engine as the alerts and the Monday brief. It splits the change into segments and ranks them by how much each one contributed. Because the cause is computed before the model sees it, two people asking the same question get the same cause, and it matches what the brief said on Monday. Annotations you added, such as a price change or a launch, come back with it, so the model can point at them.

The definitions resource and the weekly prompt

The server publishes one resource, joinsight://definitions. It's a Markdown document with every metric definition: MRR, movements, one-time revenue, active subscribers, trial conversion, churn, ARPU, LTV, at-risk MRR, channel, attributed MRR and share, revenue per visitor, GA4 sessions and Plausible visits, keyword revenue, keyword opportunities and how Stripe and RevenueCat are deduplicated. It's the same text as the metric definitions page and the dashboard tooltips. Each tool's description also includes the definitions that apply to it, so the model has them even if the client never reads the resource.

It also publishes a prompt, weekly_review. It calls get_overview for 7 days, explain_change on net new MRR, get_keyword_opportunities, get_alerts_and_health and get_revenue_leaks. Then it asks for a summary under 200 words: MRR and its change, the top causes, one suggested action (the largest leak or a keyword opportunity, whichever puts more money at stake), and any alert or data-health issue. Clients that support prompts show it as a ready-made command. It's the Monday brief, on demand, inside your assistant.

What every response contains

Each response is a JSON object with a few fields that keep answers honest:

• as_of: when the response was computed.
• currency: the workspace's reporting currency.
• freshness: one entry per source with its status, last sync time and last error, plus the snippet (active, silent or not_installed). If Stripe hasn't synced since yesterday, the model can see it and say so. Search Console lags 2 to 3 days, and the server tells the model to mention it when recent search data matters.
• modeled: true only for keyword revenue and keyword opportunities.
• attributed_share: on tools that credit revenue to pages, channels or posts. This is the share of new MRR linked to a first touch, with the customer counts and a note that the rest has an unknown source, not zero. The model is told to say how much could be attributed.
• Money: as an object with minor units and a formatted string. The model is told to quote the formatted string, so it never rounds or reformats an amount.
• Periods: with start, end and a label such as "Last 30 days".

Lists return at most 50 rows. Tools that page return a next_cursor. Pass it back as cursor to get the next page. A cursor belongs to the tool that issued it, and passing it to another tool returns an error that tells the model to call again without one.

Questions and the tools that answer them

A good first question is one you already know the answer to, such as last month's MRR. If the answer matches your billing dashboard, you have a baseline for trusting the less obvious ones.

What the model sees

Mostly aggregates and labels: totals, plan names, channel labels, page paths, search queries and post captions. Four tools return individual customers. list_customers, get_revenue_leaks and get_customer_journey include each customer's country, plan, first touch and your app's user ID. get_cancellation_reasons and get_customer_journey include the comment a customer wrote when canceling in Stripe's customer portal, in their own words. Emails are always masked to the first letter of the name and of the domain, as in j•••@e•••.com, and responses say emails_masked: true. A masked email isn't unique, so get_customer_journey refuses one and asks for the customer id instead.

The model never sees your Stripe, RevenueCat or Google credentials, your Joinsight keys, or full customer emails. Nothing is passed through to Stripe or Google: every answer comes from data Joinsight already synced. See Security, privacy and your data for what Joinsight stores in the first place.

Limits and errors

Each connection can make 60 tool calls a minute. Past that, the tool returns a message with the number of seconds to wait. A bad input, such as an invalid date or an unknown customer, returns a message the model can act on. If a query fails on Joinsight's side, the tool returns "Joinsight could not compute this right now. Try again in a minute."

When the workspace is locked because the account's trial or subscription ended, every tool call returns a message saying that access is paused and where the owner can choose a plan. The connection itself stays in place, so the same assistant answers again as soon as a plan is active, with numbers that kept syncing in the meantime. See When access ends.

Every tool call is recorded in the workspace's audit log, whether it succeeded or not, with the tool, its arguments, how long it took, and which app and user made it. Settings → AI assistants shows the latest 50 under "Recent tool calls": when, which app, which member, the tool and its arguments, whether it answered, and how long it took. Every member of the workspace can see it.

Tokens and revoking access

The connection uses OAuth 2.1 with PKCE. You never handle a token yourself, but it helps to know how they behave:

• Access tokens last 1 hour. The client renews them on its own.
• Refresh tokens last 30 days and rotate. Each renewal retires the old refresh token and issues a new pair, so a connection you use keeps working, and one left unused for 30 days lapses.
• Reuse detection. If a refresh token that was already used is presented again, Joinsight assumes it leaked and revokes the whole connection, for both the leaked copy and the real app. You then approve the app again. The same happens if an authorization code is replayed.
• Membership. If you are removed from a workspace, your connections to it stop working at once.

Tokens are stored as hashes, never in plain text.

To see and revoke connections, open Settings → AI assistants. "Connected apps" lists each connection with the app's name, a "Self-registered" badge if the app registered itself, who connected it and when, and when it was last used. "Disconnect" revokes every token of that connection immediately. Any member can disconnect the apps they connected, and owners can disconnect anyone's.

Signing out everywhere ends your browser sessions, not your MCP connections. To cut off an assistant, disconnect it under Settings → AI assistants.