Skip to Content

How do I track the prompts developers send in Cursor (beta)?

Rius traces Cursor chats (beta) through its plugin, and it records the prompts developers send once they turn on content for a folder with /rius-enable-content-here. Each chat becomes one trace in the console, with your prompts, the agent’s answers, tool calls and subagents. A folder enabled with /rius-enable-here sends structure only, with no prompt text. Cursor traces carry no token counts or cost.

How it works

The Rius plugin hooks into the Cursor agent and sends each chat to Rius as a trace. It is built for the agent in the editor and for cursor-agent, and in the beta it is tested mostly on headless cursor-agent -p runs. Tracing is off until each developer turns it on for a folder, so installing the plugin sends nothing.

The trace for a chat has this shape:

  • Session: an agent span named cursor session is the root.
  • Turn: a chain span for each turn, meaning one prompt and everything it caused.
  • Answer: a model span for the answer, with its model name.
  • Tool call: a tool span for each tool call. A shell command that exits non-zero is marked as an error.
  • Subagent: an agent span under the Task call that started it.

Whether the text of a prompt leaves the machine depends on the command a developer runs in the folder:

  • /rius-enable-here sends the span tree, timing, model names, tool names and error types. No prompts or answers.
  • /rius-enable-content-here also sends your prompts and the agent’s answers, plus tool inputs and outputs.

With content on, the plugin replaces the secrets it recognizes, such as cloud and API keys, tokens and private keys, with a marker before anything is sent, and caps each value at 32 KB. This is best effort, so a secret in a format it does not know is sent. A disabled folder (/rius-disable-here) always wins over an enabled parent.

Set it up

Cursor support needs plugin 0.6.0 or later, macOS or Linux, and Python 3.9 or later on your PATH.

  1. Install the plugin. In Cursor’s plugin settings, choose Import from Repo and enter glassflow/rius-coding-agents. The Cursor guide covers the clone option and older cursor-agent versions that ignore plugin hooks.
  2. Sign in. In the agent chat, run /rius-login, pick a workspace in the browser and click Connect.
  3. Turn prompt capture on for a folder. In the repo you want to watch, run the content command, then check it:
/rius-enable-content-here
/rius-status
  1. Work as usual. Prompts from that folder and everything under it are sent. Developers who should not send text run /rius-enable-here instead, or /rius-disable-here to stop a folder. /rius-off pauses tracing for the current chat only.
  2. Set a team limit if you need one. The environment variable RIUS_CAPTURE_CONTENT=false sends structure only from every folder, and RIUS_CURSOR_ENABLED=false turns Cursor tracing off everywhere. These settings can only turn things off.

Everyone in the workspace can read what is sent, so agree as a team which repos send content. For workspace setup across a team, the Claude Code team guide walks through it, and the same plugin covers Cursor.

What you get

  • The prompt in context. Open a chat in the trace list at console.rius-glassflow.com. Each turn is one span holding the prompt, with the answer, tool calls and subagents beneath it. Spans arrive while you work: when the chat starts, after each answer and every 20 tool calls.
  • Who sent it. From plugin 0.6.2, every span carries the email the developer signed in with, so chats appear under their name on the Users page. The Sent by column in the trace list shows whose key sent each chat. Chats sent with a key stored by use-key carry no user and land in the anonymous row.
  • An agent view. Chats show up as an agent named cursor on the Agents page. The plugin sends no heartbeat, so Cursor chats do not appear under Instances. A chat counts in the agent’s figures once its trace is closed, which happens when Cursor ends the session, or when your next chat starts after 12 idle hours.
  • Ask in plain language. The plugin brings the Rius MCP server as rius. Click Connect next to it in Cursor’s MCP settings, or run cursor-agent mcp login rius, then ask for your last failed chat or the tools your chats called most.

What Cursor does not report:

  • Tokens and cost. Cursor does not give plugins token counts, so Cursor traces show no token usage and no cost.
  • Cloud and background agents. They are not traced.
  • Tab completions. They are not traced.
  • Span times. They are when Cursor reported each step, not when the model call started.
  • Beta coverage. The plugin is tested mostly on headless cursor-agent runs. Tool calls in an interactive cursor-agent session, the agent in the editor and Import from Repo are not tested yet.

For how Cursor compares with the other agents, see Coding agents.

When something else fits

  • If you need token counts and cost per developer, Claude Code and Codex report them, and the Claude Code team guide shows usage and cost per developer.
  • If prompt text should never leave a machine, use /rius-enable-here for structure only, which still gives you the span tree, timing, models and tool names.

Still stuck? Get help

Last updated on
Quickstart