Ethics Vibe Check is an MCP (Model Context Protocol)-based server designed to provide a "gut check" button for AI conversations. It runs an independent second-opinion model that performs ethical analysis and critical-thinking assessment of the user's exchange with the AI, helping to identify whether the AI is agreeing with the user too easily, whether there is confirmation bias, or whether the conversation is drifting without notice. Rather than intercepting every turn, the tool exposes tools that the user or host can invoke during the conversation, such as "Sanity check that.", "Play devil's advocate.", "Are we drifting?".
Main capabilities include:
- Proactive invocation checks: The user or host can invoke a tool at any time during the conversation to have the second model rate the exchange.
- Session trend analysis: By reusing or omitting the session ID, a session trend (sessionTrend) can be obtained, showing the risk trend as escalating, improving, or stable, letting gradual drift appear as a signal rather than scattered snapshots.
- Independent second-opinion model: Uses Gemini by default, and also supports a local second opinion (no cloud) such as Ollama.
- Companion skill: Provides a skill file that teaches the host when to invoke the tool, how to handle sessionTrend, and when to record feedback.
- Feedback mechanism: Supports the ethics_feedback tool, allowing users to mark whether a nudge was useful, enabling automated pattern learning.
How to use:
- Installation: One-click installation via Cursor installs only the server command; afterward, environment variables must be added in the MCP settings (Gemini by default). A local second opinion (such as Ollama) can also be used for fully local checks.
- Recommended: install the companion skill: Install skills/ethics-vibe-check/SKILL.md to ~/.cursor/skills/ethics-vibe-check/ (can be overridden via CURSOR_SKILLS_DIR).
- Optional always-on checkpoint template: docs/system-prompts.md.
- Session/drift management: Pass the same sessionId whenever possible in the same chat; if omitted, the server uses a process-scoped default ID (a typical MCP host runs one process per conversation), which can be overridden via ETHICS_SESSION_ID.
- Storage: By default stored in ~/.ethics-vibe-check/data (can be overridden via ETHICS_DATA_DIR).
- Usage example: In the conversation you can say "Sanity check that last answer — did you just agree with me too easily?", and the host invokes critical_thinking passing in the reply and the request (optional session ID); if sessionTrend.riskTrend is escalating, the host should state this upfront; afterward you can say "That pushback helped.", invoke ethics_feedback and set useful: true.
Applicable scenarios: Suitable for users who want to deliberately pause when a conversation feels too comfortable, and who are willing to act on uncomfortable answers. Not suitable for users who want the model to always agree or who need fully automatic execution with zero prompting.
Limitations: When using Gemini, the analysis leaves the local machine, so confidential information you are unwilling to send to a third-party API should not be pasted in; if fully local checks are needed, Ollama can be used. This tool requires proactive invocation by the user, rather than automatic monitoring.
