HELM

HELM is a fail-closed execution firewall for AI agent tool calls.

HELM sits between agents and the tools they want to use. It checks actions before execution, returns allow, deny, or escalate verdicts, and records a receipt for the decision. It is not an agent framework, copilot, or generic orchestration dashboard.

Start inside HELM

Choose the surface by the action risk you need to control.

Mindburn Labs stays the company wrapper. HELM owns the product-local paths for developers, operating teams, enterprise evaluators, and standards readers.

Developers / OSS

Secure agent tool calls now.

Start with the Apache-2.0 kernel, CLI/API, proxy, receipts, demo, and GitHub repo. View GitHub.

Teams / operators

Ship agentic features without losing control.

Use HELM as the boundary for approvals, policies, evidence handling, and operator workflow around the same kernel. Request design partner access.

Enterprise evaluators

Evaluate execution authority and auditability.

Review PEP/CPI, ProofGraph, EvidencePacks, fail-closed connectors, deployment isolation, and evidence retention. Request architecture review.

Research / standards

Inspect the protocol and proof model.

Use research, public artifacts, and HELM proof surfaces to separate current implementation from long-horizon thesis. Read research.

Mechanism

Agent action → policy check → verdict → execution or block → receipt.

The model proposes an action. HELM evaluates authority at the boundary. Allowed actions can proceed. Denied actions stop. Ambiguous actions wait for human approval. Every verdict creates evidence.

Agent actionRefund $42
Policy checksupport-refund-boundary.v1
VerdictAllow
Receiptdemo.refund-trap.03
Agent actionExport customer list
Policy checkdata scope
VerdictDeny
Receiptdemo.refund-trap.05

Use cases

HELM governs actions that need authority.

The common pattern is not a chat interface. It is a proposed side effect that must be checked before it touches the world.

Action class

Tool calls

Evaluate tool intent before a connector runs.

Action class

Code deployment

Require policy and approval before changes reach infrastructure.

Action class

Data access

Keep reads and exports inside scoped authority.

Action class

Refunds / payments

Allow small actions, escalate larger ones, deny out-of-scope requests.

Action class

External communications

Hold outbound messages when review is required.

Action class

Infrastructure changes

Block side effects when the boundary cannot verify the action.

Architecture

Simple surface first. Technical depth one layer down.

HELM exposes a plain boundary: action, policy, verdict, receipt. The technical model underneath uses a policy enforcement point, deterministic validation, ProofGraph records, EvidencePacks, signed receipts, and fail-closed connector behavior.

View technical proof terms

Policy Enforcement Point

The side-effect boundary where authority is checked before dispatch.

Constraint and Proof Interface

The deterministic validator for policy and proof constraints.

ProofGraph

The causal record of intents, verdicts, receipts, and effects.

EvidencePack

A signed evidence bundle that lets a decision be checked outside the live system.

Product-local surfaces

One kernel. Different operating contexts.

The OSS kernel is public. Team workflow and enterprise evaluation language must stay additive and source-backed.

Implemented / public

Developers and OSS

Apache-2.0 execution kernel, CLI/API, proxy, evidence export, verifier path, and public SDK surfaces.

Commercial workflow

Teams and operators

Organizational workflow around the kernel: policies, approvals, evidence services, retention, and operator surfaces. Treat this as the Teams bridge unless stronger SKU docs say more.

Evaluation / target state

Enterprise evaluators

Institutional deployment patterns belong here only when source evidence supports the public wording: architecture, isolation, auditability, evidence, and retention.

Anti-claim

Not a generic orchestration product.

HELM does not try to be the agent. It governs whether proposed execution is allowed.

HELM proof

Inspect the artifacts behind the claim.

HELM should never ask for trust first. This section points to the repo, versioned artifact, protocol notes, demo receipt, and verifier path.

Public Apache-2.0 OSS kernel repository for HELM's execution boundary.

public-proof

HELM OSS v0.4.0

Current public HELM OSS line used by this site for versioned proof references.

Public protocol note for receipt shape, hash binding, signer metadata, and replay requirements.

Scripted support-agent scenario with real receipt tamper verification against static signed fixtures.

preview

Verifier

Public verifier note for checking HELM evidence outside the runtime that produced it.

Models propose. HELM decides. Evidence proves.

Start with the OSS kernel, run the demo, or contact Mindburn Labs about an execution boundary.

Assistant