Correctover MCP Server

What is this?

Correctover is the first MCP server that verifies AI outputs in real-time.

While every other MCP server connects your AI tools to data sources, Correctover sits in the execution path and ensures every LLM response is correct, complete, and reliable — before it reaches your editor.

Your AI Tool (Cursor/Claude Desktop/Windsurf)
        │
        ▼
┌─────────────────────────────────┐
│  Correctover MCP Server         │
│                                 │
│  ① Route → picks best provider  │
│  ② Execute → calls LLM API     │
│  ③ Verify → 6-dim check        │  ← This is what nobody else does
│  ④ Heal → auto-fix or failover  │
│  ⑤ Deliver → verified output    │
│                                 │
└─────────────────────────────────┘
        │
        ▼
  LLM Providers (OpenAI / Anthropic / DeepSeek / ...)

Why you need this

AI APIs don't just fail with HTTP 500. The worst failures are silent:

• Response looks valid but contains hallucinated data
• JSON output is truncated mid-object
• Provider silently degrades output quality over time
• Token usage spikes without warning

Correctover catches all of these. Every response passes through 6-dimension validation:

If validation fails, Correctover automatically retries or fails over to another provider — and validates again. This is not simple retry. This is verified failover.

Failover ≠ Correctover. Failover switches providers. Correctover switches and verifies the output is correct before delivering it.

MCP Protocol Compatibility

This server implements the Model Context Protocol specification version 2025-11-25, using JSON-RPC 2.0 over stdio transport.

The protocol layer uses an adapter pattern — adding new transport types (WebSocket, gRPC) in the future will not affect the core validation engine. We track MCP specification updates closely and test compatibility on every protocol version release.

Supported features:

• ✅ JSON-RPC 2.0 over stdio
• ✅ initialize / tools/list / tools/call / notifications
• ✅ Multi-tool support (chat, verify, providers, health)
• 🔜 WebSocket transport (planned)
• 🔜 Streaming tool results (planned)

Installation

One-line JSON config

Add to your MCP client config (e.g., ~/.cursor/mcp.json):

{
  "mcpServers": {
    "correctover": {
      "command": "npx",
      "args": ["-y", "correctover-mcp-server"],
      "env": {
        "OPENAI_API_KEY": "sk-...",
        "DEEPSEEK_API_KEY": "sk-...",
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

That's it. No servers to deploy. No dependencies to install. No configuration files to manage.

Build from source

git clone https://github.com/Correctover/mcp-server.git
cd mcp-server
go build -o correctover-mcp-server .

# Then in your MCP config:
# "command": "/path/to/correctover-mcp-server"

Supported Providers

Configure providers via environment variables. Only configured providers are active.

Proxy/Mirror support: Each provider's base URL can be overridden via {PROVIDER}_BASE_URL environment variable. Perfect for self-hosted proxies, API gateways, or regional mirrors (e.g. OPENAI_BASE_URL=https://your-proxy.com/v1).

BYOK (Bring Your Own Key): Your API keys stay on your machine. Correctover connects directly to providers — no proxy, no middleman, no data leakage.

Tools

chat

Send a chat message with automatic verification and self-healing.

Parameters:

• messages (required): Conversation messages in OpenAI format
• model: Model name or "auto" for automatic selection
• provider: Force a specific provider
• temperature: Sampling temperature
• max_tokens: Maximum response tokens
• system_prompt: System prompt to prepend

Returns: The LLM response + a validation report showing which dimensions passed/failed.

health

Check which providers are active and ready.

providers

List all supported providers with configuration details.

stats

Show session statistics: total calls, validation pass rate, failover count.

Example Output

Every chat call returns a validation report:

╔══════════════════════════════════════╗
║   Correctover Validation Report     ║
╠══════════════════════════════════════╣
║ Provider: deepseek                  ║
║ Latency:  847ms                     ║
║ Model:    deepseek-chat             ║
║ Score:    6/6                       ║
║ Passed:   true                      ║
╠══════════════════════════════════════╣
║ ✅ structure  PASS                   ║
║ ✅ schema     PASS                   ║
║ ✅ latency    PASS                   ║
║ ✅ cost       PASS                   ║
║ ✅ identity   PASS                   ║
║ ✅ integrity  PASS                   ║
╠══════════════════════════════════════╣
║ ✓ All dimensions passed              ║
╚══════════════════════════════════════╝

How it works

1. Route — Selects the best available provider based on priority and health
2. Execute — Sends the request to the selected provider
3. Verify — Validates the response across 6 dimensions
4. Heal — If validation fails: auto-retries with same provider, or fails over to next provider, then re-validates
5. Deliver — Returns the verified response with a full validation report

This is the MAPE-K control loop (Monitor-Analyze-Plan-Execute-Knowledge) applied to LLM API reliability, running in real-time at sub-millisecond decision overhead.

Who is this for?

• Developers who use Cursor/Claude Desktop and want more reliable AI responses
• Teams building AI-powered applications who need output guarantees
• Enterprises in regulated industries (finance, legal, healthcare) where AI output errors have real consequences
• Anyone tired of silently wrong AI outputs breaking their workflow

FAQ

Q: How is this different from LiteLLM / OpenRouter? A: They route requests. We route + verify outputs. Think of it as the difference between a delivery service and a delivery service with quality inspection.

Q: Do you store my API keys? A: No. Keys stay on your machine. We connect directly to providers. Zero proxy, zero data collection.

Q: Does this work with Cursor? A: Yes. Add the JSON config above to ~/.cursor/mcp.json and restart Cursor. Done.

Q: What if I only have one provider? A: Still works. You get 6-dimension validation on every response. Failover kicks in when you add more providers later.

Sponsor

If Correctover saves you from a silent AI failure, consider supporting:

• ☕ $5/month — Thank you + priority issue responses
• 🚀 $29/month — Private Discord + monthly update briefings
• 🏢 $99/month — Enterprise sponsor, logo on README

→ Sponsor on GitHub

Need Help Integrating?

For team deployments, custom validation rules, or dedicated support:

📧 hello@correctover.com

License

Apache-2.0