MCP Feedback Enhanced

Interactive feedback collection for AI assistants in Cursor, with real-time pending message injection and feedback enforcement.

• Rich Feedback UI: Bottom panel with chat bubbles (AI left, user right), markdown support, image input (paste/drag-drop/file picker), and quick replies.
• Cursor Hooks Integration: Pending user messages are injected into the agent loop in real time via preToolUse hooks, using HTTP-based delivery from the extension's in-memory store.
• Feedback Enforcement: Periodically re-injects rules into the agent's context to prevent them from being overlooked during long sessions.
• Pending Queue: Queue messages and images while the agent is busy — they are delivered at the next tool call via the preToolUse hook. Delivered pending messages appear as user bubbles with a 📤 pending badge.
• Auto-Configuration: Automatically sets up ~/.cursor/mcp.json, ~/.cursor/hooks.json, and ~/.cursor/rules/mcp-feedback-enhanced.mdc on activation.
• Auto-Focus: Bottom panel activates automatically on extension startup and when the agent requests feedback.
• Image Support: Paste (Cmd+V), drag-drop, or file picker for images. Images are displayed in chat, included in pending messages, and passed to the LLM via MCP image responses.
• Secure & Local: All data stays on your machine.

Install MCP Feedback Enhanced from Open VSX or build from source.

The extension auto-configures on activation:

• ~/.cursor/mcp.json — MCP server registration
• ~/.cursor/hooks.json — preToolUse hook (pending message delivery + rules refresh)
• ~/.cursor/rules/mcp-feedback-enhanced.mdc — always-applied rule for interactive_feedback usage

1. The AI Agent calls interactive_feedback with a summary.
2. The MCP Feedback panel activates automatically at the bottom.
3. Type your feedback, attach images, or click a Quick Reply.
4. The AI receives your input and proceeds.

Pending Messages: Submit a message anytime via the panel. If the agent is busy, the message is queued and injected at the next tool call via the preToolUse hook.

Agents sometimes forget to call interactive_feedback before ending their turn. This wastes user requests (the user must type a new message instead of responding for free via the MCP panel).

Enforcement triggers:

• Count-based: After maxToolCalls (default: 15) tool calls without interactive_feedback
• Time-based: After maxMinutes (default: 5) minutes without interactive_feedback

When triggered, the hook denies one tool call and silently re-injects rules into the agent's context. Counter resets after — the retry passes through immediately.

Override defaults via ~/.config/mcp-feedback-enhanced/enforcement-config.json (maxToolCalls, maxMinutes).

┌───────────────┐     WebSocket     ┌──────────────────┐     stdio     ┌─────────────┐
│  Webview Panel │◄────────────────►│  Extension (Hub)  │◄────────────►│  MCP Server  │
│  (Bottom)      │                  │  wsHub.ts         │              │  mcp-server/ │
└───────────────┘                   └──────────────────┘               └─────────────┘
                                      │ HTTP endpoints │
                                      │ /pending       │
                                      │ /health        │
                                      └───────┬────────┘
                                              │ HTTP GET
                                       ┌──────▼──────┐
                                       │ Cursor Hooks │
                                       │  consume-pending.js │ → pending + rules refresh
                                       │  hook-utils.js      │ → shared utilities
                                       └─────────────┘

~/.config/mcp-feedback-enhanced/
├── servers/<hash>.json              # Running extension instances
├── feedback-state.json              # Tool call counter for enforcement
├── enforcement-config.json          # Optional enforcement config overrides
└── logs/hooks.log                   # Hook debug log

DeepMem gives your AI persistent memory across sessions and IDEs. When paired with MCP Feedback Enhanced, noteworthy interactions are automatically distilled and stored as memories.

How it works: After every interactive_feedback call, a post-feedback hook buffers the interaction. Every 5 non-trivial interactions, an LLM extracts decisions, preferences, gotchas, and patterns, then stores them in the running DeepMem daemon via HTTP.

Requirements: DeepMem daemon running (bun run start or deepmem start) with ~/.deepmem/daemon.json present. Optional: ~/.deepmem/deepmem.yaml with completion.apiUrl / completion.apiKey / completion.model for LLM distillation.

• "Connecting...": Reload the window (Cmd+R / Ctrl+R).
• MCP Server Error: Check ~/.cursor/mcp.json is configured correctly.
• Hooks Not Working: Verify ~/.cursor/hooks.json contains entries with _source: "mcp-feedback-enhanced".
• Panel Not Focusing: Use Command Palette → "MCP Feedback".

• Install: Open VSX or install via VSIX
• Repository: GitHub

Powered by Model Context Protocol