Overwatch
Overwatch is a runtime security proxy for MCP (Model Context Protocol) servers. It intercepts tool calls, detects Tool Shadowing attacks (CVE-2025-6514), and enforces policies before execution.Why Overwatch?
AI agents using MCP servers have access to powerful tools - databases, filesystems, and APIs. Without access control:- A single prompt injection can drop database tables
- Malicious MCP servers can shadow trusted tools
- Dangerous operations execute without human review
- No audit trail of what the AI agent did
Quick Start
Key Features
Tool Shadowing Detection
Detects CVE-2025-6514 where malicious MCP servers impersonate trusted tools:- Schema hashing - SHA-256 fingerprints of tool definitions
- Collision detection - Same tool name across multiple servers
- Mutation monitoring - Schema changes mid-session
- Description analysis - 30+ suspicious patterns
Policy Engine
Configure fine-grained access control:Session Grants
After approving an operation, grant temporary permissions:| Grant | Duration | Use Case |
|---|---|---|
| Once | Single use | One-time operations |
| 5 min | 5 minutes | Short tasks |
| 15 min | 15 minutes | Moderate tasks |
| Session | 24 hours | Full development session |
overwatch sessions revoke.
Audit Logging
Complete audit trail of all operations:Reliability Features
- Circuit breaker - Protects against upstream failures
- Request timeouts - Prevents hanging operations
- Connection recovery - Exponential backoff reconnection
- Fail modes - Configure behavior when upstream fails