Skip to content
Docs

Connect an assistant

Ask for predictions in a conversation. Claude and ChatGPT connect to Causilo through one MCP server address, on the web or from the command line, and you sign in with your Causilo account once.
On this page
MCP server addressTEXT
https://mcp.nums.world/mcp

How connecting works

The assistant opens a Causilo consent page. Sign in with the code we email you, check the workspace, and choose Allow. Causilo creates a key named after the assistant, such as Claude connection, in that workspace. The assistant never sees the key; it holds a token that only the Causilo MCP server accepts.

AccountWhat it decides
Your assistant accountWhere the conversation happens. Any Claude or ChatGPT account works.
The Causilo account you sign in withThe organization and workspace that get the key, and whose allowance the predictions use.

An address that is not yet a member of an organization gets a personal organization on its first sign-in.

Whether the assistant asks before it calls a tool depends on the assistant and its settings. Only a prediction uses your allowance: a predict_table call, or a file prediction when its command runs. The other tools are free.

On the web

Claude

  1. In claude.ai, open Plugins, then Add, then Add custom connector.
  2. Name it Causilo and enter https://mcp.nums.world/mcp.
  3. Choose Connect and allow access on the Causilo consent page.

ChatGPT

  1. In chatgpt.com, open Plugins, then Add, then Add custom MCP server.
  2. Name it Causilo, enter https://mcp.nums.world/mcp as the server URL and keep OAuth as the authentication.
  3. Check I understand and want to continue, choose Create as a plugin, then choose Connect and allow access on the Causilo consent page.

From the command line

Add the server, then sign in. Your browser opens the same consent page and returns to the command line when you allow access.

Claude Code

Add the serverBASH
claude mcp add --transport http causilo https://mcp.nums.world/mcp

Then run /mcp, select causilo and choose Authenticate.

Codex

Add the serverBASH
codex mcp add causilo --url https://mcp.nums.world/mcp

Codex opens the sign-in when the server is added. Run codex mcp login causilo to sign in again.

Ask for a prediction

Give the assistant a table and say which column to predict. In ChatGPT, attach a CSV, TSV, Excel or Parquet file. Claude sends the table as rows in the request, up to 1,000 rows. Claude Code and Codex do the same for a small table and send a larger one as a file, as described below. For a larger table anywhere else, call the API. Say regression or classification when the target holds whole numbers; otherwise the server chooses from the target and reports its choice.

Example requestTEXT
Use the Causilo connector to predict y. Regression.
Context rows: x=1,y=10; x=2,y=20; x=3,y=30;
              x=4,y=40; x=5,y=50; x=6,y=60.
Query rows: x=7 and x=8.

The answer holds the prediction for every query row, up to 1,000 values a call. For a larger result, predict from a file in Claude Code or Codex, or call the API. The MCP server passes the table to the API and keeps neither the table nor the predictions.

ToolWhat it does
predict_tablePredicts one column and returns the predictions.
estimate_tableChecks whether a table fits and roughly how long it takes.
inspect_tableShows the columns, types, missing values and first rows.
usage, limitsShow the allowance and the request limits.
get_upload_commandGives a shell command that predicts from a file. Claude Code and Codex.

Predict from a file

In Claude Code and Codex, name the file and the column to predict. The assistant asks Causilo for a command and runs it in your shell. The command sends the file for one prediction and saves the predicted rows with their predictions as a CSV file at the path you name. When you give no separate query file, the rows with an empty target are predicted, and the first column, row_number, tells where each row is in your file, counting from 0.

Example requestTEXT
Use Causilo to predict the empty rows of the late column in orders.csv.
Save the predictions to late_predictions.csv.
ItemValue
FilesCSV, TSV, Excel or Parquet, up to 100 MiB each
ResultUp to 2,000,000 predicted values and 64 MiB. A larger one is refused before anything is charged.
CommandGood for 10 minutes and one prediction
AllowanceUsed once, when the command runs the prediction. Getting the command is free.
ConnectionOne made by signing in. A connection that sends an API key gets no command.
Shellbash, zsh or another POSIX shell with curl

The result file changes only when the command succeeds. If the command stops after it has sent the file, the prediction may have been charged without a result. The assistant is told to ask you before it gets a new command. The MCP server holds the file and the result in memory during the request and stores neither.