HELM AI Kernel
Why AI Agents Need an Execution Firewall
Open-source execution kernel, CLI, MCP, conformance, verification, and compatibility.AI agents can propose useful work, but tool calls are where proposals become side effects. A model can ask to read a ticket, draft a reply, call an MCP tool, export customer data, or run a shell command. Those requests need a deterministic execution boundary before anything reaches the tool.
HELM AI Kernel is an open-source execution firewall for MCP and AI agents. It intercepts proposed tool calls, evaluates policy before dispatch, records ALLOW, DENY, or ESCALATE decisions, and emits signed receipts that can be verified offline.
Execution Firewall Flow
flowchart LR
Agent["AI agent"] --> Proposal["Proposed side effect"]
Proposal --> Kernel["HELM AI Kernel"]
Kernel --> Policy["Policy and schema evaluation"]
Policy --> Verdict["ALLOW, DENY, or ESCALATE"]
Verdict --> Dispatch["Dispatch only when allowed"]
Verdict --> Receipt["Signed receipt"]
Receipt --> Verifier["Offline verification"]Mermaid source
flowchart LR
Agent["AI agent"] --> Proposal["Proposed side effect"]
Proposal --> Kernel["HELM AI Kernel"]
Kernel --> Policy["Policy and schema evaluation"]
Policy --> Verdict["ALLOW, DENY, or ESCALATE"]
Verdict --> Dispatch["Dispatch only when allowed"]
Verdict --> Receipt["Signed receipt"]
Receipt --> Verifier["Offline verification"]The key idea is simple: the agent can propose, but HELM decides whether the side effect is authorized. Unknown MCP servers and tools fail closed before fixture dispatch. Schema-pinned calls can be allowed. A DENY decision produces a receipt, and a flipped-verdict copy fails verification.
Run the local proof path without an account or production credentials:
git clone https://github.com/Mindburn-Labs/helm-ai-kernel.git
cd helm-ai-kernel
make build
bash scripts/launch/demo-mcp.sh
bash scripts/launch/demo-proof.sh
Star the repo if you want to follow the MCP execution-firewall roadmap: https://github.com/Mindburn-Labs/helm-ai-kernel