Table of contents

Integrating our AI agent with Cursor / Claude Code

<aside> 💡

With the release of v0.1.0 of our VSCode / Cursor extension, you can now integrate our agent as an mcp into both Cursor and Claude Code!

</aside>

1. Open VSCode / Cursor
2. Sign-in to NCProf - Use ****NCProf: Sign in from the command palette to sign in. This will redirect you to the browser.
3. Setup integration of your choice -
   1. NCProf: Setup Cursor Integration - This will automatically update your system’s Cursor config to add the nCompass Chat MCP
   2. NCProf: Setup Claude Code Integration - This will automatically update your system’s claude code config to add the nCompass Chat MCP
4. Check that the MCP is available to your agent
5. Continue vibe coding!
   1. If you “@”-add an .ncu-rep / .nsys-rep / .json.gz / .pftrace to your code context, Claude Code / Cursor will automatically know to invoke the nCompass agent to run analysis on these file types.

Interacting with systems trace data (nsys / torch.profile) using the AI assistant

Opening the AI Assistant View

Open the AI assistant by selecting the NCProf Assistant: Focus on chat view from the Command Palette

Add a trace file to the chat context to chat with it

Use the + or @ functionality to point the chat to a trace file. This could be a .nsys-rep or a .json.gz file and it allows you to bring a trace into the context and chat with it.

<aside> 💡

The agent we’ve released in v0.0.20 of our extension can add an arbitrarily sized trace to your chat context and let you ask very specific and detailed questions about the trace.

</aside>

<aside> ⚠️

If you run into an auth error when using the agent - you have to sign in. Do this by running NCProf: Sign in from the Command Palette and follow the instructions.

</aside>

https://www.loom.com/share/9de1e0c8eece4d65a46874497defe244

Video legend:

0:00 Introduction to AI agent

1:00 Example Question

2:08 Answer to question 1 - what should I care about in this trace

4:18 Answer to question 2 - how do kernels map to operations from LLMs

Link to trace from video : Use https://docs.ncompass.tech/collaborate-on-trace-files to open the link with our extension.

Link to full response from AI agent

Interacting with GPU kernels trace data (ncu) using the AI assistant

Opening the AI Assistant View

Open the AI assistant by selecting the NCProf Assistant: Focus on chat view from the Command Palette

Add a trace file to the chat context to chat with it

Use the + or @ functionality to point the chat to a trace file. This could be a .nsys-rep or a .json.gz file and it allows you to bring a trace into the context and chat with it.

<aside> 💡

This hasn’t been released on our public version of the extension yet. It will be released in version v0.1.0

</aside>

<aside> ⚠️

If you run into an auth error when using the agent - you have to sign in. Do this by running NCProf: Sign in from the Command Palette and follow the instructions.

</aside>

https://www.loom.com/share/5bb95731f8294bb39084adae3ca1571b

Video Legend

0:00 Introduction to kernel side features

0:41 Agent demo