prashar32 / riskkernel

What is riskkernel?

RiskKernel is a self-hosted reliability layer for AI agents written in Go. It sits in front of your agents and enforces hard limits on cost, token usage, loop iterations, and wall-clock time. The core promise: your agent proposes, but deterministic Go code decides whether the action proceeds. It works as a transparent proxy you point your existing app at with a single environment variable, or you can use the Python SDK for deeper control. Every call is metered, priced, and budget-enforced. Runs that exceed their limits are killed cleanly with HTTP 402, and you can resume crashed runs from the last checkpoint.

Why is it gaining traction?

The hook is simple: runaway agents are expensive. A loop that burns $400 overnight while you sleep is a real problem nobody's shipping a clean solution for. RiskKernel differentiates by being genuinely self-hosted with zero telemetry, deterministic enforcement (no LLM deciding when to stop), and a crash-resume mechanism that makes long-running agents practical. The human approval gates are another draw -- you can pause side-effecting tool calls (shell execution, file writes, deployments) until a human approves or denies. It speaks OpenAI and Anthropic APIs natively, and exports OpenTelemetry GenAI spans to your existing observability stack.

Who should use this?

Teams running AI agents in production who need cost predictability and operational safety nets. Backend engineers building agentic workflows who can't afford surprise bills. DevOps teams that need audit trails and kill switches for autonomous agents. If you're using LangGraph, CrewAI, or AutoGen and wish they shipped guardrails, this fills that gap. Early-stage but worth evaluating if governance and cost control are non-negotiable for your use case.

Verdict

RiskKernel solves a real pain point with a clean architecture and zero-friction adoption path. At 10 stars with a 0.899% credibility score, it's clearly early-stage and one-person maintained -- test coverage looks solid but the project needs more battle-testing in production environments before I'd trust it with a high-stakes agent. The proxy-onramp is compelling enough to try in staging. Watch the roadmap and consider contributing if the mission resonates.