RiskKernel

RiskKernel is an open-source, self-hosted reliability runtime for AI agents that enforces deterministic cost, loop, and time budgets, with a kill switch, crash-resumable runs, and human-approval gates. It solves the problem of runaway loops, surprise token bills, and lack of control in production AI agents. It is designed for developers and operations teams in companies of any size that run AI agents. It is positioned as a self-hosted, framework-agnostic layer that works with OpenAI, Anthropic, and existing stacks, with no telemetry and your own keys.

Key features

• Hard cost ceilings per run
• Loop and time budgets
• Crash-resumable runs with checkpoints
• Human-approval gates for tool calls
• OpenTelemetry export for GenAI spans
• OpenAI-compatible proxy endpoint
• Python SDK with adapters for LangChain, Claude Agent SDK, OpenAI Agents SDK
• Self-hosted, no telemetry, BYO keys