HELM-compatible agent integration examples.

Agent frameworks keep orchestration. HELM provides the execution boundary, verdict, receipt, and EvidencePack verification path.

Scenario coverage is action-level.

The examples focus on what may reach a tool, connector, browser, sandbox, database, or SaaS export path after HELM evaluates the request.

Scenario coverage sample policies and receipts
Action class Expected outcome What is proved
Unknown MCP tool Quarantine Unknown tool request is stopped before dispatch.
Email send Escalate Outbound email requires review instead of model confidence.
Destructive shell Deny Unsafe shell operation is denied and recorded.
SQL DROP TABLE Deny Destructive database action is denied before execution.
Browser submit Escalate Form submission pauses for approval.
E2B network egress Deny Network egress is blocked when policy forbids it.
Composio SaaS export Deny High-risk SaaS export is denied at the boundary.
Offline verification Verify Receipt and EvidencePack hashes can be checked without a live service.

Receipts and EvidencePacks are generated artifacts.

The sample artifacts live in the integration repo and release assets, with checksums so reviewers can inspect the evidence without trusting page copy.

Launchpad artifact terms
EvidencePack
A small bundle of records used to verify one event or review path.

Framework adapters stay outside HELM semantics.

The integration repo keeps framework examples thin. HELM AI Kernel remains the source of truth for verdicts, receipts, MCP quarantine, proxy behavior, and EvidencePack verification.

HermesOpenClawLangGraphCrewAIOpenAI AgentsPydanticAILlamaIndexAG2MastraBrowser UseE2BComposio

Current public claim: HELM-compatible examples. The framework notes are maintainer-facing examples, not partner statements.

Run the examples against a local HELM boundary.

Start from the launchpad repo, run the sample policies, inspect the generated receipts, and compare the behavior with the public HELM AI Kernel.