Correctover — MCP Runtime Security

The runtime security layer for the MCP ecosystem.

We build the runtime security layer for AI agent ecosystems. Correctover enforces security conformance on every MCP tool call — validating LLM outputs, blocking injection attacks, preventing credential hijacking, and auto-recovering from failures at 22μs P50 latency.

What We Do

Related MCP server: Kluster.ai Verify MCP

Real Data — Not Benchmarks

We don't simulate. We collect real API responses from production MCP servers and verify them.

20,000 Verified API Traces

Collection methodology:

• 43.6 minutes of continuous collection at 13.7 API calls/second

• 68.88% conformance rate (13,776 conformant / 6,224 non-conformant)

• 30% fault injection rate for stress testing

• Every record includes: request, response, latency, validation result, fault classification

Third-party independent verification: 120,426 conformance re-calculations by @babyblueviper1 — full consistency confirmed.

The complete 20K dataset is in ./data/ — download, verify, fork, do whatever you want. No gatekeeping.

Research & Publications

CCS v1.0 — Runtime Security Conformance Standard

The first formal conformance standard defining how agent runtimes should validate tool execution results at runtime.

• Paper: CCS Standard v1.0 Final (PDF)

• DOI: 10.5281/zenodo.21234580

• Protocol Spec: RFC-001 | RFC-002

Key findings from 20K real traces:

• P50 validation latency: 22μs

• Self-heal rate: 97.4% (engine auto-retries/fails over on failed validation, then re-validates)

• Rule coverage: 88 detection rules (64 high-confidence)

• 561 distinct fault variants cataloged across all major LLM providers

Fault Taxonomy

We maintain a living fault taxonomy derived from real-world MCP server failures:

• 215 distinct fault types classified across 7 severity levels

• 19 CVE-class vulnerabilities identified across MCP implementations

• Categories: RCE, SSRF, cloud credential hijacking, path traversal, output injection, privilege escalation

Upstream PR Contributions

We don't just report — we fix. Our contributions go directly into major agent frameworks:

Community Validation

Real researchers using our work in production:

MCP Server — Product

The runtime verification engine packaged as an MCP server for your AI tools.

Quick Start

npm install -g correctover-mcp-server

Add to your mcp.json (Cursor, Claude Desktop, Windsurf):

{
  "mcpServers": {
    "correctover": {
      "command": "correctover-mcp-server",
      "env": {
        "OPENAI_API_KEY": "sk-...",
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

BYOK — your keys stay on your machine. No proxy, no data collection.

How It Works

Your AI Tool (Cursor / Claude Desktop / Windsurf)
        │
        ▼
┌─────────────────────────────┐
│   Correctover MCP Server    │
│   ┌───────────────────────┐ │
│   │  6-Dim Validator      │ │
│   │  ├─ Structure         │ │
│   │  ├─ Schema            │ │
│   │  ├─ Latency           │ │
│   │  ├─ Cost              │ │
│   │  ├─ Identity          │ │
│   │  └─ Integrity         │ │
│   └───────────────────────┘ │
│          │                  │
│   ┌──────▼──────┐           │
│   │ Failover    │           │
│   │ Engine      │           │
│   └──────┬──────┘           │
└──────────┼──────────────────┘
           │
    ┌──────┼──────┬──────────┐
    ▼      ▼      ▼          ▼
 OpenAI  Anthropic  DeepSeek  Qwen ...

Supported Providers

Tools

Ecosystem Adoption

Real download numbers from public package registries (last 30 days):

All organic growth — no paid promotion. CCS standard package seeing highest growth rate (tens of thousands percent increase from baseline).

CCS SDK — Agent Governance

Python SDK for embedding governance into agent frameworks. Fail-closed by design.

pip install correctover-ccs

from ccs import govern

@govern(policy="default")
def my_tool(args: dict) -> str:
    return "result"

# Governance evaluates BEFORE function runs
# If denied → PermissionError, function never executes

Supported frameworks: CrewAI, AutoGen, LangGraph/LangChain

Observer hooks (default):  governance_crash → tool EXECUTES ❌
CCS decorators (ours):     governance_crash → tool BLOCKED ✅

Security Audit Reports

We publish detailed security audits of MCP server implementations:

→ mcp-security-audits

Methodology: Source code analysis → fault injection → runtime verification → CVE classification

Findings to date: 506 security findings across 3 major repositories, 5 vulnerability types confirmed cross-repo.

Ecosystem & Links

Contact

Security reports: wangguigui@correctover.com
BD / Enterprise: wangguigui@correctover.com
GitHub: @Correctover

License

Apache 2.0 © Correctover