Verdict MCP

An unbypassable MCP gatekeeper that enforces code completeness, test coverage, and premium UI/UX standards on autonomous AI agents.

The Problem

AI agents suffer from Context Drift, Illusion of Completion (Hallucination), and Quality Degradation:

• Mark tasks DONE but leave empty placeholders

• Skip critical error handling

• Generate unstyled "developer art" UIs

• Write superficial tests

Related MCP server: mcp-llm-eval

The Solution

Verdict sits between the AI agent and the file system as a programmatic gatekeeper. The agent cannot mark a task as complete until this server audits the code, validates design standards, and verifies 95%+ test coverage with 80%+ mutation score.

Quick Start

pip install -r requirements.txt

Create a plan.md in your project root:

# Project Plan
## TASK_001: Setup authentication
- [ ] Create auth_handler.py
- Status: PENDING

Then run the server:

python -m verdict_mcp

MCP Primitives

Resources

Tools

Prompts

Validation Lifecycle

[Agent Writes Code]
       ⬇
[submit_task_for_audit]  → AST verification (pass/TODO rejected)
       ⬇
[enforce_ui_standards]   → Premium UI validation (glassmorphism enforced)
       ⬇
[run_strict_test_suite]  → pytest + 95% coverage + 80% mutation score
       ⬇
[Task → COMPLETED]

Features

• AST-level auditing — detects pass, # TODO, missing exception handling, missing docstrings

• UI dictatorship — rejects unstyled components, enforces design tokens (glassmorphism, neon)

• 95% coverage gate with mutation testing — tests must catch injected bugs

• Git-backed snapshots — auto-snapshot before audit, auto-rollback on test failure

• Dependency DAG — tasks cannot start until dependencies are complete

• Plan hash chain — SHA-256 integrity check prevents plan tampering

• Coverage caching — 30-second TTL avoids redundant test runs

• Structured output — tools return typed data (not just strings) for better agent integration