Add Cursor Agent adapter

[jackjin1997]

Context

Cursor is the most popular AI-powered IDE. Adding a Cursor agent adapter would attract massive community interest and provide a compelling IDE-vs-CLI comparison.

Approach

Cursor doesn't have a direct CLI mode, but we could:

1. Use Cursor's background agent API (if available)
2. Or wrap the Cursor composer via automation

Looking for community input on the best approach.

Why This Matters

Cursor vs Claude Code vs Gemini CLI would be the most requested comparison in the AI coding space right now.

[IlyasFardaouix]

I think the key issue here is that we're trying to integrate with a non-CLI API (Cursor's background agent API) without much information on how to use it. Can you provide more details on the API's available methods and any authentication requirements?

[jackjin1997]

Agreed — Cursor vs Claude Code would be a highly requested comparison.

The honest status: Cursor doesn't currently expose a public CLI or headless API for agent-mode execution. The "background agent" feature runs inside the IDE and there's no documented way to invoke it programmatically from outside.

Options we're considering:

1. Wait for Cursor CLI — if Cursor ships a CLI mode (similar to how Claude Code works), we'll add it immediately
2. Manual benchmark — run tasks manually in Cursor and record results. Less reproducible but still valuable for comparison
3. Cursor extension scripting — if someone finds a way to script the Cursor agent via its extension API

For now, we're focusing on agents with proper CLI support (Claude Code, Gemini CLI, Codex CLI, Aider). Once we have 4+ agents with reproducible Docker-based results, adding Cursor manually would still be a good data point.

If anyone has experience with Cursor's internal APIs, please share!

[IlyasFardaouix]

While we're waiting for more details on Cursor's background agent API, I'd like to explore option 3: Wrap the Cursor composer via automation. We could leverage tools like Selenium or PyAutoGUI to simulate user interactions with the IDE and programmatically execute tasks. This approach would sidestep the lack of a direct API, but we'd need to carefully consider the reliability and performance implications of automating the UI. Can we discuss the feasibility of this approach and potential use cases?