Connect Your CAT Mock Log to Claude: What It Can Do

You can connect your Karma Yogi account to Claude and ask questions about your own CAT prep in plain language, such as "which of my mocks are due for revision" or "how have my DILR scores moved." The connector exposes 13 tools: six that read, four that log new records, three that correct existing ones, and none that delete. A generic chatbot cannot do this, because it has no access to your log. A study tracker without a connector cannot do it either, because it has no conversation. The combination is the point: your numbers stay in one place, and the analysis happens where you can ask follow-up questions.

This post covers the setup, the tools, one real example, and where the limits are. The reference page with every tool and example prompt is the connector setup guide.

What it is, in one paragraph

MCP, the Model Context Protocol, is a standard way for an AI assistant to call tools exposed by another service. Karma Yogi runs a remote MCP server. After you approve a one-time sign-in, Claude can call that server's tools on your behalf, scoped to your account only. It cannot see anyone else's data, and the access token it receives cannot be used against the regular Karma Yogi app API.

Setup in five steps

1. In claude.ai, open Settings and go to the custom connectors section under Connectors.
2. Add a custom connector with the server URL https://karma-yogi-api.onrender.com/mcp.
3. claude.ai registers itself with Karma Yogi automatically, so there is no client ID or secret to type.
4. Sign in with your existing Karma Yogi account, or approve the connection if you are already signed in.
5. You are returned to claude.ai, connected.

The flow only works when started from Claude. Opening the Karma Yogi connect link directly shows a dead-end message by design. To disconnect, go to Profile, then Preferences, then Account and Data, then Connected apps. Access tokens are short-lived (15 minutes by default), so access stops within about 15 minutes of disconnecting.

The 13 tools

Three details matter for CAT specifically. First, mock sections come from your own subjects, so a CAT log has VARC, DILR and Quant, while a GATE or JEE log has its own. Second, the logging tools will not duplicate a mock with the same name and date, which matters when you repeat a request. Third, edits send only the fields that change: asking Claude to fix one Quant score leaves your VARC and DILR notes alone.

What it did with my log

This week I pointed Claude at my own data. It read my goals (CAT, 29 November 2026, a 50-hour weekly target), my study statistics (659 sessions, 38,437 minutes, 197 active days, a 58-minute average session, 190 minutes at the longest), and my full-mock list. That list held 40 mocks, 39 of them scored, going back to 10 May.

The most useful output was not a chart. It was a comparison I had not made myself: raw scores of 59 to 65 had returned percentiles from 66.91 to 98.46 depending on whether the paper was a real past CAT or a coaching mock. I wrote that up with the full table in CAT Past Papers vs SIMCAT: The Same Score, Two Percentiles. The arithmetic behind it, means, ranges and the controlled comparison from August onward, was done by a script on the data the connector returned, not by eyeballing. I then checked each figure against the log before publishing.

That last step is the part to copy. The connector gives Claude your real numbers, which removes the biggest failure of asking a chatbot about your prep: it invents plausible ones. It does not remove the need to check a calculation before you act on it.

Prompts worth trying

• "Analyse my last 20 mocks and name my weakest section by accuracy, not by score."
• "Which of my mocks are due for a revision round, and how overdue is each one?" The spaced schedule runs R1 three days after the mock, then 7, 14, 21, 30, 45 and 60 days, each counted from when the previous round was actually done. The tool descriptions spell this out so Claude does not compute it as test date plus a fixed offset, which would overstate how overdue you are.
• "Compare my past-paper mocks against my coaching mocks. Are the percentiles comparable?"
• "I just finished a 45-minute Quant session on geometry, log it."
• "My SIMCAT 14 Quant score was 33, not 30. Fix it."

Limits you should know

• Rate cap: 30 tool calls per minute and 400 per hour per account. It exists to stop a runaway loop, and a normal conversation does not approach it.
• Cold starts: the first request after a long idle period can take 10 to 30 seconds while the API wakes up. Retry rather than assuming it failed.
• Large lists get truncated by the client. My 40-mock list was big enough that the raw result had to be saved to a file and processed with a script. Ask for a recent window (the last 20) if your client struggles.
• No access to AI Insights, the in-app mentor chat or consolidated mock notes. Those are not exposed as tools.
• No edits to subjects, goals or preferences. Those shape how everything else is interpreted, so they stay in the app.
• What you type to Claude is governed by Anthropic's privacy policy, not Karma Yogi's. Karma Yogi does not see your conversations.

Who this is for, and who it is not

It is for someone with at least a few weeks of logs, because a model reading five sessions has nothing to find. It is for people who already ask "what should I fix" after each mock and want a faster second opinion than a spreadsheet pivot table. It is not a substitute for judgement: Claude will read a trend confidently even when n is eleven, so ask it for the sample size every time.

Karma Yogi is free and works for CAT and any other exam you can define subjects for.