A "Limitless Refill" Spell for Cursor, Specifically Targeting Quota Anxiety This article is excerpted from the WeChat public account: AI Apprentice. Cursor's request quota has always been a pain point for many users. 500 requests per month may sound generous, but when tackling tasks like debugging or implementing complex features, a few rounds of dialogue can quickly deplete the quota. As user Ryo lamented, spending 145 yuan only to burn through 0.3 yuan per query feels painfully wasteful. Then, user ultrasev shared an intriguing technique—or rather, a "spell." He added this prompt: "... Every time you think you've fixed it, ask me in the terminal if I'm satisfied (read -P 'Repair completed. Are you satisfied? (y/n) ' response && echo $response). Only exit after receiving 'yes.' If I don't reply or reply otherwise, continue repairing, repeating the process until you read 'yes.' This is critically important." This cleverly leverages Cursor's tool calling feature. Effectively, it stretches one request into multiple interactions. Instead of completing the task silently, the AI pauses at each step via the terminal tool to ask, "Done. Happy with this?" Unless you reply 'y,' it keeps refining. Even for moderately complex tasks, this allows you to iterate within a single request quota until the issue is fully resolved. A brilliant little cost-saving trick. Now, someone has turned this idea into an MCP tool. Each chat with Cursor counts as one request. But often, fulfilling a single requirement takes multiple back-and-forths, turning what should be one charge into n charges—hardly economical. Enter mcp-feedback-enhanced MCP server. Find the open-source repo here: https://github.com/Minidoracat/mcp-feedback-enhanced In short, it consolidates multiple interactions into one request. You submit a request, Cursor works on it, then uses this MCP to solicit feedback. You provide input, it adjusts, all within a single dialogue—no extra charges. Supported platforms: Cursor | Cline | Windsurf | Augment | Trae Demo of the Tool in Action Let's test its effectiveness. We'll make a normal request or explicitly state: "I need to test mcp-feedback-enhanced MCP." After addressing the first requirement, it invokes the MCP, opening a GUI window for text input. Upon receiving feedback, it processes the response, then reopens the window—iterating as needed. This way, one request fuels multiple exchanges. Observe the extended dialogue below. Note: Cursor Agent caps tool calls at 25 per session. Beyond that, a new request is triggered. I recall writing prompts that make AI "think actively." Often, a single dialogue exhausts all 25 tool calls. GUI Interface Overview Here's how the MCP operates, focusing on this GUI: The summary merely displays Cursor's dialogue recap—ignore it. Similarly, the command panel is informational. Settings handle layout preferences—nothing critical. Thus, the MCP is straightforward to use. One caveat: It supports both web and QT interfaces. If the QT interface malfunctions, prompt the Agent to fix it. MCP Installation Guide Now, let's install this MCP. Past articles covered various MCPs—some installations are trickier. This one involves a GUI. Win Configuration Complete setup (Windows is the most challenging) First, Windows setup—the toughest platform. Follow the same approach as before: Review the mcp-feedback-enhanced README (on GitHub) alongside my earlier guide: Windows MCP Error Savior: Fix Cursor Config Issues in 1 Minute.. The official docs recommend UV installation. When my Cursor Agent failed to install UV, it defaulted to the local Python interpreter: C:\Python312\python.exe. Final config file: For Mac/Linux, follow the official docs. If UV installation stumps you, feed the docs to Cursor Agent. Optimizing Continuous Calls To maximize MCP's iterative potential, tweak these prompts: # MCP Interactive Feedback Rules 1. During any process, task, or dialogue—whether querying, replying, or completing a milestone—always invoke MCP mcp-feedback-enhanced. 2. Upon receiving user feedback (if non-empty), recall MCP mcp-feedback-enhanced and adapt behavior accordingly. 3. Only cease MCP calls when the user explicitly states "end" or "no further interaction needed," marking process completion. 4. Unless terminated, all steps must repeatedly invoke MCP mcp-feedback-enhanced. 5. Before task completion, solicit feedback via MCP mcp-feedback-enhanced. Add to settings rules: Or project rules: Configuration complete.
