RiskKernel

Uncapped agent runs are a production liability: no ceiling on spend, no deterministic exit on loops, no gate before a destructive tool call fires. RiskKernel addresses this by acting as a metered proxy and enforcement runtime between your code and your LLM provider. Point your existing agent at the local endpoint with one environment variable, and every API call is measured against the dollar, token, loop count, and wall-clock budgets you set. The kill switch fires mid-loop the moment a ceiling is crossed — before the spend lands, according to the vendor — and a checkpoint written at each step means a crashed run resumes without re-executing paid work.

The differentiating feature is determinism. The vendor makes an explicit architectural choice: budgets, gates, and kill switches live in statically-typed Go code, not in prompts or framework configuration. That means the enforcement logic runs identically every time regardless of what the model reasons. The human-approval gate — which blocks side-effecting tool calls and routes them for review before execution — inherits that property. The LLM proposes; compiled code decides whether the action proceeds.

The runtime is Apache 2.0 and self-hosted with no telemetry; provider keys are never stored in plaintext and nothing phones home, which is verifiable in the source. OpenTelemetry export emits GenAI spans for cost, loops, and checkpoints to Datadog, Grafana, Honeycomb, or any compatible backend. SDK adapters cover LangChain, the Claude Agent SDK, and the OpenAI Agents SDK. Where RiskKernel breaks down: the hosted dashboard is in private beta, so teams that need a visual run-management UI are on a waitlist. Organisations with LLM frameworks outside the three listed adapters will need to handle integration themselves at the SDK layer.