Blind-Auditor-MCP
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Blind-Auditor-MCP audit this Python function for security issues and style compliance"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
🛡️ Blind Auditor - MCP Server
"Code -> audit -> amend -> audit again -> pass."
Blind Auditor is a mandatory code auditing system built on the MCP (Model Context Protocol). It uses a unique "Thinking Isolation" mechanism to force AI Agents to enter an independent "audit phase" and self-review their code before outputting the final result.
🧠 Core Philosophy: Thinking Isolation
Traditional AI coding is often "generate and output," which allows errors and biases to slip through. Blind Auditor introduces a middle layer:
Intercept: When the Agent wants to output code, it must first submit it to Blind Auditor.
Isolate: Blind Auditor does not return the result immediately. Instead, it injects a mandatory system instruction, forcing the Agent to pause its current persona and switch to a "Ruthless Auditor" role.
Audit: In this isolated context, the Agent must scan the generated code line by line against the predefined
rules.json.Release: The code is unlocked and returned to the user only when the audit score meets the threshold (default > 80) and there are no Critical issues.
Related MCP server: MeshMind
🎯 Key Features
🛡️ Zero Trust Architecture: Default distrust of the Agent's initial draft; it must pass an audit.
💰 Zero Extra Cost: Reuses the host IDE's current session model, requiring no additional API Key.
⚖️ Bias Removal: Forces a perspective switch via Prompt injection to break generation inertia.
📏 Strict Compliance: Hard-codes team code standards (
rules.json) into the generation process, which is more effective than simple Prompts.🔄 Auto-Fix Loop: Automatically triggers a "fix-resubmit" loop upon audit failure until compliance is met or the maximum retry count is reached.
🚀 Quick Start
1. Prerequisites
This project requires Python 3.10+ and uv.
# Install uv (if not already installed)
curl -LsSf https://astral.sh/uv/install.sh | sh
# Enter project directory
cd blind-auditor
# Install dependencies (automatically creates virtual environment)
uv sync2. Configure Audit Rules (rules.json)
Create or edit rules.json in the project root. This is the auditor's "Rulebook".
{
"project_name": "MyCoreProject",
"strict_mode": true,
"max_retries": 3,
"rules": [
{
"id": "SEC-001",
"severity": "CRITICAL",
"description": "Strictly prohibit hardcoding any API Keys, passwords, or sensitive Tokens.",
"weight": 50
},
{
"id": "PERF-001",
"severity": "WARNING",
"description": "Avoid database queries or IO operations inside loops.",
"weight": 15
},
{
"id": "STYLE-001",
"severity": "PREFERENCE",
"description": "Use snake_case for variable names.",
"weight": 5
}
]
}Configuration Fields:
severity:CRITICAL: Fatal error. If found, the audit fails immediately. Usually carries a high penalty.WARNING: Warning. Moderate penalty; accumulation can lead to failure.PREFERENCE: Style preference. Low penalty.
weight: The penalty points deducted when this rule is violated (Total score: 100).
3. Configure IDE (Antigravity / Cursor / Windsurf)
Step A: Add System Prompt (Critical)
To ensure the Agent knows it must use this tool, add the following instruction to your .cursorrules or IDE System Prompt settings:
# Blind Auditor Protocol
**CRITICAL INSTRUCTION**:
You are integrated with the **Blind Auditor MCP**.
Whenever you generate code for a user request, you **MUST NOT** output the code block directly to the user.
**Execution Protocol**:
1. **Draft**: Generate the code internally in your thought process.
2. **Submit**: Call the `submit_draft` tool with your code and language.
3. **Obey**: The tool will return a strict "Audit Instruction". You MUST follow it exactly to audit your own code.
4. **Finalize**: Only when the tool returns "✅ AUDIT PASSED" can you display the final code to the user.
**Prohibited**:
- Do NOT output code blocks before passing the audit.
- Do NOT skip the audit step for "simple" changes.Step B: Configure MCP Server
Add the following to your MCP client configuration file (e.g., claude_desktop_config.json or IDE settings):
Note: Replace /path/to/your/blind-auditor with the actual absolute path where you cloned this repository.
{
"mcpServers": {
"blind-auditor": {
"command": "uv",
"args": ["run", "--directory", "/path/to/your/blind-auditor", "blind-auditor"]
}
}
}🔧 Tool Details
1. submit_draft
Submit a code draft.
Input:
code(content),language(programming language)Behavior: Locks the session and returns mandatory audit instructions.
2. submit_audit_result
Submit your audit conclusion.
Input:
passed(bool): Whether you believe it passed.issues(list): List of issues found.score(int): Score from 0-100.
Behavior:
If
score < 80, forcespassed=False.If passed, unlocks the code.
If failed, increments retry count and requires the Agent to fix and resubmit.
3. reset_session
Resets the state and clears the retry count.
🔁 Workflow Diagram
graph TD
User["User Request"] --> Agent
Agent["Agent Generates Draft"] -->|1. submit_draft| MCP
MCP -->|2. Inject Audit Instructions| Agent
subgraph Isolation ["Thinking Isolation"]
Agent -->|3. Self-Review| Agent
Agent -->|4. submit_audit_result| MCP
end
MCP -->|5. Verdict| Decision{"Passed?"}
Decision -->|No - Issues Found| Retry["Retry Count +1"]
Retry -->|Limit Not Reached| Fix["Agent Fixes Code"]
Fix -->|Resubmit| Agent
Decision -->|Yes - Score >= 80| Final["✅ Output Final Code"]
Retry -->|Limit Reached| Force["⚠️ Force Output - With Warning"]❓ Troubleshooting
Q: The Agent always outputs code directly without calling tools. A: Check if the System Prompt is configured correctly. You must explicitly tell the Agent "Do NOT output code directly". You can also manually remind it in the chat: "Please audit via Blind Auditor first".
Q: Why does it fail even if I give the code 100 points?
A: Check if any CRITICAL rules in rules.json were triggered. Current logic mainly relies on the score passed by the Agent, but if passed is True while score < 80, the system will force a rejection.
Q: Which programming languages are supported?
A: Theoretically, all languages are supported. Blind Auditor itself does not parse code syntax but relies on the Agent's understanding to match descriptions in rules.json.
🛠️ Development Guide
# Run server
uv run blind-auditor
# Or run directly with Python module
uv run python -m src.main
# Debug mode (output to stderr)
# View print statements in src/main.py📄 License
MIT License
Available Tools
4 toolsreset_sessionB
Reset the current audit session.
Inputs: no named parameters. Returns: result
submit_audit_resultD
Submit the audit result.
Inputs: passed (required), issues (required), score. Returns: result
submit_draftC
Submit a code draft for audit.
Inputs: code (required), language. Returns: result
update_rulesA
Update audit rules configuration.
Args:
action: Operation to perform - "add", "remove", "update", or "list"
rule_id: Rule identifier (required for add/remove/update)
severity: Rule severity level - "CRITICAL", "WARNING", or "PREFERENCE" (for add/update)
description: Rule description (for add/update)
weight: Point deduction weight 0-100 (for add/update)
Returns:
Status message with operation result
Examples:
# List all rules
update_rules(action="list")
# Add a new rule
update_rules(
action="add",
rule_id="SEC-001",
severity="CRITICAL",
description="No hardcoded API keys",
weight=50
)
# Remove a rule
update_rules(action="remove", rule_id="SEC-001")
# Update a rule
update_rules(
action="update",
rule_id="SEC-001",
description="Updated description"
)
Inputs: action (required), rule_id, severity, description, +1 more. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 4 tools
Each tool has a clearly distinct purpose with no overlap: reset_session handles session management, submit_audit_result and submit_draft handle different submission types (final results vs. code drafts), and update_rules handles rule configuration. The tools target different resources and actions, making misselection unlikely.
Three tools follow a consistent verb_noun pattern (reset_session, submit_audit_result, submit_draft), while update_rules also fits this pattern but uses a different verb style ('update' vs. 'reset'/'submit'). The naming is mostly consistent with only minor deviation in verb choice.
With 4 tools, this is well-scoped for an audit server. Each tool earns its place by covering core audit workflow stages: session management, rule configuration, draft submission, and result submission. The count is neither too sparse nor bloated for the domain.
The toolset covers the essential audit lifecycle: configure rules, submit drafts, submit results, and manage sessions. A minor gap exists in lacking a tool to retrieve or review audit results/history, but agents can work around this using the existing submission tools. Core workflows are well-supported.
Maintenance
Related MCP Connectors
Hosted YouTube MCP server for Claude, ChatGPT and Cursor, with each action switched per account.
An MCP server that gives your AI access to the source code and docs of all public github repos
Cloudflare Workers MCP server: ai-agent-scratchpad
MCP server teaching AI agents to implement TideCloak: auth, E2EE, IGA, security analysis
Related MCP Servers
- AlicenseAqualityBmaintenanceA local-first MCP server that gives AI coding agents persistent memory and controlled commands. Features a git-backed markdown knowledge vault with FTS5 search, surgical section edits, token-aware context budgeting, and a sandboxed command engine with human approval gates. Works with Claude Code, Cursor, Copilot, Gemini, and more.621Apache 2.0
- AlicenseAqualityCmaintenanceA single MCP server that merges three context-engineering ideas into one toolset for AI agents76 npm1MIT
- AlicenseNot gradedqualityDmaintenanceA local MCP server that gives AI coding agents persistent memory and context across sessions.6 npmMIT
- AlicenseNot gradedqualityCmaintenanceA sandbox-first MCP server that enables AI agents to safely perform reverse-engineering tasks in fully isolated Docker containers, featuring zero-trust security, declarative YAML-based skills, and white-hat ethical safeguards.MIT