OpenAI-Compatible Execution Boundary

If your applications already speak the OpenAI chat-completions API, you can put a policy boundary in front of them without changing a line of client code. HELM AI Kernel ships a proxy that accepts OpenAI-compatible requests, evaluates them against your policy, and only then forwards to the upstream model.

Requests that violate policy never reach the provider: they return a DENY or ESCALATE verdict instead, and the decision is recorded as a signed receipt either way. Allowed traffic flows through with receipt metadata attached, so every model call in your fleet becomes attributable and replayable.

Start it with one command (make proxy in the repository, or the demo script below), point your existing client's base URL at it, and watch the receipt stream record what your agents actually ask for.

Gateway Policy Path

flowchart LR
    Client["OpenAI-compatible client"] --> Proxy["HELM proxy"]
    Proxy --> Policy["Policy and receipt boundary"]
    Policy --> Upstream["Upstream model when allowed"]
    Policy --> Denial["DENY or ESCALATE"]
    Upstream --> Receipt["Receipt metadata"]
    Denial --> Receipt

flowchart LR
    Client["OpenAI-compatible client"] --> Proxy["HELM proxy"]
    Proxy --> Policy["Policy and receipt boundary"]
    Policy --> Upstream["Upstream model when allowed"]
    Policy --> Denial["DENY or ESCALATE"]
    Upstream --> Receipt["Receipt metadata"]
    Denial --> Receipt

git clone https://github.com/Mindburn-Labs/helm-ai-kernel.git
cd helm-ai-kernel
make build
bash scripts/launch/demo-openai-proxy.sh

Source Truth

• Quickstart
• Execution security model
• MCP integration
• Verification