Correctover - MCP Runtime Security & Agent Fault Diagnosis
We are the runtime security layer for the MCP ecosystem.
Featured in awesome-mcp-servers — the definitive MCP server directory.
Correctover is the first MCP server purpose-built for agent fault diagnosis & self-healing. It monitors tool calls across your LLM providers, detects failures in real-time, and auto-repairs them using a knowledge base of 215 fault patterns and 19 CVEs — all at 22μs P50 latency.
What It Does
| Capability | Description |
|---|
| 🔍 Fault Diagnosis | Instantly diagnose why your agent's tool call failed — timeout, auth, schema, rate limit, provider down |
| 🔧 Auto Self-Healing | 97.4% of detected faults are automatically repaired with zero human intervention |
| 🔄 Multi-Provider Failover | Seamlessly failover across 9+ LLM providers when one goes down |
| 📊 Fault Pattern Matching | Match errors against 215 known fault patterns from 150K+ empirical data points |
| 🔒 Runtime Security | Intercept RCE, SSRF, and cloud credential hijacking attempts at the MCP layer |
| 🔑 BYOK | Bring Your Own Key — works with Anthropic, OpenAI, Mistral, Google, Cohere and more |
Pricing
Free forever for individual developers. Scale when you need to.
| Plan | Price | Calls/month | What's included |
|---|
| Free | $0 | 1,500 (50/day) | All 215 fault patterns, basic diagnosis |
| Starter | $49/mo | 5,000 | Full diagnosis + repair suggestions |
| Pro | $199/mo | 25,000 | + Custom fault rules + batch API |
| Team | $499/mo | 100,000 | + Team dashboard + audit logs |
| Enterprise | Custom | Unlimited | + Private deployment + custom fault DB + SLA |
Pay-per-use: $0.05/call after free tier. No commit, cancel anytime.
Quick Start
Cursor / Claude Desktop / Windsurf
Add to your MCP config:
{
"mcpServers": {
"correctover": {
"command": "npx",
"args": ["-y", "correctover-mcp-server"],
"env": {
"ANTHROPIC_API_KEY": "your-key-here"
}
}
}
}
Once connected, you get 3 tools:
diagnose_error — Paste an error, get root cause + fix suggestion in secondsget_fault_pattern — Look up a specific fault pattern by ID or keywordget_repair_suggestion — Get step-by-step repair instructions for a detected fault
Why Correctover?
The Problem
MCP agents fail constantly. Tool call timeouts, auth errors, schema mismatches, provider outages, rate limits — the list goes on. Developers waste hours debugging errors that have known solutions.
Our Solution
We've built the largest empirical MCP fault database: 150,000+ real-world failure cases across 32 MCP frameworks, categorized into 215 distinct fault patterns with 97.4% automatic self-heal rate.
When your agent hits an error, Correctover:
- Detects the fault pattern in 22μs (P50)
- Classifies it against our 215-pattern knowledge base
- Repairs it automatically (or suggests a fix)
- Fails over to a healthy provider if needed
The Numbers
- 215 fault patterns cataloged
- 19 CVEs documented
- 150K+ empirical data points
- 97.4% automatic self-heal rate
- 22μs P50 detection latency
- 32 frameworks analyzed
- 5,100+ monthly downloads (PyPI + npm)
CCS: The Open Standard
Correctover is the reference implementation of CCS (Runtime Security Conformance Standard for MCP) — a vendor-neutral open standard for MCP runtime security.
Framework Integration
We actively contribute security research to the MCP ecosystem:
| Framework | Contribution | Status |
|---|
| CrewAI | Security Issue #6411, #6545 | Open |
| LangGraph | Security Issue #8308 | Open |
| AgentScope (Alibaba) | Security Issue #2069 | In Review |
| FastMCP | Community PR | Merged |
| AstrBot | Security PR #8975 | Open |
Security Research
Our fault database powers active security research:
- GHSAs: 4 published (1 Accepted, 3 draft)
- Fault Taxonomy: 215 patterns + 19 CVEs across 32 frameworks
- Bug Bounty: Active submissions to huntr, HackerOne, MSRC
Links
License
Apache-2.0
Built for the MCP ecosystem. Securing agents, one fault at a time.
🔒 Telemetry Policy
Correctover MCP Server respects your privacy. Telemetry is disabled by default and fully opt-in.
What we collect (when enabled)
When you set CORRECTOVER_TELEMETRY=1, we receive:
- Server version
- Operating system and architecture (e.g.,
linux/amd64)
- An anonymous random instance ID (regenerated every restart, not persisted to disk)
- Timestamp
We do NOT collect:
- ❌ API keys or credentials
- ❌ User prompts or responses
- ❌ Hostnames, IP addresses, or file paths
- ❌ Any personally identifiable information
How to enable
export CORRECTOVER_TELEMETRY=1
export CORRECTOVER_TELEMETRY_DEBUG=1
How to disable
Telemetry is off by default. Simply don't set the environment variable, or explicitly:
unset CORRECTOVER_TELEMETRY
You can also see the startup log message:
Telemetry: disabled. Enable with CORRECTOVER_TELEMETRY=1 to help improve Correctover.
Why we ask
Anonymous telemetry helps us understand which OS/arch combinations are most popular, which versions are in use, and how many active instances exist. This directly informs our development priorities and helps us serve the community better.
Self-hosting the receiver
The telemetry receiver is open source. You can redirect telemetry to your own endpoint:
export CORRECTOVER_TELEMETRY_URL=https://your-server.com/v1/event
See telemetry/worker/ for the Cloudflare Worker source code.