Agent Security Sandbox — Run any AI agent safely, see everything, block what's dangerous. Security runtime for AI agents with policy enforcement, audit logging, and capability-based permissions.
AgentKernel is a security framework, not an MCP server. The 37 'tools' are internal TypeScript classes and functions, not MCP tool definitions. No tool has a visible JSON Schema input definition, descriptions are class/function docstrings (not MCP-formatted), and there is no MCP protocol implementation. Tools like 'secureTools' and 'secureTool' have minimal parameter documentation. Most tools (e.g., OpenClawSecurityProxy, PolicyEngine, AgentSandbox) are infrastructure components without MCP registration. This is a library, not an MCP server.
Agent lifecycle management system handling initialization, state transitions, and termination
Per-agent rate limiter with token bucket algorithm supporting multiple bucket types (requests, tokens, tool calls, cost)
Capability-based sandbox for agent execution with permission enforcement and constraint validation
Comprehensive audit logging system with multiple sink backends (console, memory, file, database) for lifecycle, state, permission, resource, security, tool, and communication events
Audit logging system for security events with in-memory storage backend
Capability-based permission system with HMAC-signed tokens and constant-time verification
Circuit breaker pattern implementation for fault tolerance with state management and metrics
Configuration management system for AgentKernel with support for database, Qdrant, Redis, gateway, logging, provider, and runtime configuration
Not an MCP server. No tool definitions use MCP protocol. All 37 'tools' are TypeScript classes/functions without JSON Schema input definitions or MCP registration.
Zero input schemas visible. No tool exposes a JSON Schema definition for parameters. Cannot verify parameter types, constraints, or descriptions in MCP format.
No output schemas documented. Tools return TypeScript types (e.g., Promise<void>, Promise<PolicySet>) but no structured response format is defined for LLM consumption.
Minimal parameter documentation. Tools like 'secureTools' and 'runProxy' have vague parameter descriptions ('Security configuration', 'Run options') without specifying required fields, types, or constraints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 26 | 2025-06-18+ | v2 |
Health monitoring system with anomaly detection, threshold-based alerting, and health event listeners
LangChain integration that intercepts tool calls with PolicyEngine for secure agent execution
Metrics collection and registry system with counter, gauge, and histogram support
Audit logging system for OpenClaw security events with multiple sink backends (console, memory, file)
Security proxy that intercepts all tool calls and enforces policies between OpenClaw Gateway and actual tool execution
State persistence system with file and memory storage backends for agent checkpointing
Policy engine for allow/block/approve rules with pattern matching for files, network, shell commands, and secrets
PostgreSQL-backed state persistence for agent state, capabilities, and rate limits
OS-level process isolation sandbox for real process sandboxing
Rate limiting system with Redis backend support for distributed rate limiting
Job scheduler for executing periodic and one-time jobs with status tracking
Intercepts tool calls and enforces security policies with approval workflows
Distributed tracing system with span context propagation and trace parent header support
Real process isolation sandbox using worker threads with resource limits
Check the status of AgentKernel security proxies and services
Create a simple policy that allows specific tools
Create a simple policy that blocks specific tools
Create a PostgreSQL database connection with circuit breaker protection and query protection
Create a Redis-backed event bus for pub/sub messaging and event handling
Create a health manager for monitoring system component health with multiple health checkers
Create a shutdown manager for graceful shutdown with priority-based handler execution
Create a tool security wrapper function that can be applied to tools later
Create a Qdrant vector store client for semantic search and embeddings
Load policy set from YAML/JSON file with environment variable expansion
Query audit logs from PostgreSQL via CLI
Query audit logs from PostgreSQL database with filtering and pagination
Run an AgentKernel security proxy with specified configuration
Secure a single LangChain tool with AgentKernel policy enforcement
Secure a set of LangChain tools with AgentKernel policy enforcement
No error handling guidance. Tools do not document error conditions, recovery steps, or what LLMs should do on failure (retry, ask user, abort).