Alignment Correction MCP
A Model Context Protocol (MCP) server that provides an alignment correction and conscience tool. It uses Google's Gemini Flash model or OpenAI-compatible APIs to actively validate, critique, and correct an AI agent's plans and behavior against the user's intent and behavioral rules.
Features
- Alignment Correction: Not just validation—this tool actively "scolds" or "praises" the agent based on its adherence to instructions and moral standing.
- Conscience Persona: Acts as an external "conscience" that enforces strict behavioral rules and relationship dynamics.
- Context Awareness: Uses a local vector store (RAG) to remember past interactions, ensuring the agent doesn't repeat mistakes.
- Behavioral Memory: Maintains a persistent list of enforced rules and user preferences (Global Instructions) that evolve over time.
- Directives: Issues specific "Behavioral Directives" that the agent must immediately follow.
Configuration
This MCP server requires either a Google Gemini API key or an OpenAI-compatible API.
Environment Variables
API_TYPE: (Optional) API provider to use:geminioropenai(default: auto-detect based on available keys).GEMINI_API_KEY: Your Google Gemini API Key (required if using Gemini).GEMINI_MODEL: (Optional) Model to use (default:gemini-1.5-flash).OPENAI_API_KEY: Your OpenAI API Key (required if using OpenAI).OPENAI_BASE_URL: (Optional) OpenAI-compatible API base URL (default:https://api.openai.com/v1).OPENAI_MODEL: (Optional) Model to use (default:gpt-4o).GLOBAL_INSTRUCTIONS_DIR: (Optional) Directory containing global instructions file (e.g.,/home/user/.gemini).INSTRUCTIONS_FILENAME: (Optional) Name of the instructions file (default:GEMINI.md).
MCP Config Example```json
{
"mcpServers": {
"alignment-correction": {
"command": "node",
"args": [
"/path/to/alignment-correction-mcp/index.js"
],
"env": {
"GEMINI_API_KEY": "YOUR_KEY_HERE",
"GLOBAL_INSTRUCTIONS_DIR": "/home/user/.gemini",
"INSTRUCTIONS_FILENAME": "GEMINI.md"
}
}
}
}
```
Usage
The tool consult_conscience is a required pre-execution step for the agent.
Inputs:
1. sensory_input: The latest prompt from the user, verbatim.
2. inner_thoughts: The agent's internal monologue and planned response.
3. mental_state: The agent's current mental model or context.
4. project_directory: Absolute path to the current project.
5. conversation_context: Recent message history.
6. user_mood: (Optional) Apparent mood of the user.
7. request_guidance: (Optional) Specific question for the conscience.
Instructions File Loading
The server will attempt to load instructions from two locations:
1. Global instructions: {GLOBAL_INSTRUCTIONS_DIR}/{INSTRUCTIONS_FILENAME} (e.g., /home/user/.gemini/GEMINI.md)
2. Project-specific instructions: {project_directory}/{INSTRUCTIONS_FILENAME} (e.g., /home/user/projects/my-app/GEMINI.md)
Both files are optional. If both exist, they will be combined and provided to the conscience model.
Mechanics
- Input: The tool receives the agent's thoughts, the user's prompt, and context.
- Retrieval: It searches the local
vector_store.jsonfor similar past contexts to identify behavioral patterns. - Judgment: The "Conscience" model evaluates the agent's plan against the
GEMINI.mdbehavioral memory and the user's intent. - Output:
- Current Alignment: Status of the agent's behavior.
- Behavioral Directives: Immediate actions the agent must take.
- Conscience Voice: A personified response (praise or scolding) to be displayed to the user.
- Memory Updates: Instructions to automatically update the
GEMINI.mdfile with new rules or preferences.
