CLI

Use helm-ai-kernel to run the local proof path, connect agent clients, and inspect receipts.

Operator Front Door

On an interactive TTY, bare helm-ai-kernel (or tui / ui / dashboard) opens the full-screen operator TUI. Escape hatches keep the text catalog:

HELM_NO_TUI=1 helm-ai-kernel
TERM=dumb helm-ai-kernel
helm-ai-kernel help --all
helm-ai-kernel help --json

Press ? inside the TUI for the keyboard map. The catalog ranks Doctor, Watch, Policy, Freeze, and Threat before setup convenience. Ceremony decisions require typing APPROVE or DENY; a click never decides. See CLI I/O Convention.

First Proof

helm-ai-kernel mcp proof --json --out ~/.helm-ai-kernel/proofs
helm-ai-kernel verify --bundle ~/.helm-ai-kernel/proofs/<run-id>/evidencepacks/<run-id> --profile dev-local --allow-self-attested --json

The explicit opt-in accepts the locally generated seal as proof of internal consistency, not provenance.

Local Agent Setup

helm-ai-kernel
helm-ai-kernel setup status --format json
helm-ai-kernel setup status cursor --format json
helm-ai-kernel setup claude-code --dry-run --format json
helm-ai-kernel setup repair claude-code --dry-run
helm-ai-kernel setup remove claude-code --dry-run
helm-ai-kernel setup --client cursor --print-config

Inspect first. Apply, repair, and remove require --dry-run or explicit --yes / typed APPROVE. Cursor/VS Code status names the documented config path and never claims the editor loaded HELM. Windsurf is print-config-only.

Setup JSON reports client_state and a projected lifecycle (absent / planned / pending / configured / active / degraded / repairable). Only native_loaded is an active claim; Cursor/Windsurf/VS Code never report native_loaded.

Setup writes local client configuration and draft policy artifacts. It does not approve tools. Hermes setup writes a fail-closed pre_tool_call shell hook only; it does not write MCP. DeepSeek setup writes a Kernel hook file and points the stock DSH dsh-hooks-claude-code bridge configPath at it; it does not add a HELM-native agent runtime and does not claim npx @deepseek-ai/dsh web sees DENY.

MCP Authorization Commands

Use these commands to inspect the fail-closed MCP boundary:

Command	Purpose
helm-ai-kernel mcp authorize-call --server-id <id> --tool-name <tool>	Evaluate one MCP tool call before dispatch.
helm-ai-kernel mcp approve --server-id <id> --tools <csv> --ttl 15m --reason <text>	Returns unavailable until credential verification is configured; it does not create approval authority.
helm-ai-kernel mcp revoke --server-id <id> --reason <text>	Revoke an existing local MCP registry record.
helm-ai-kernel mcp pending --json	List servers or tools awaiting credential-verified approval.
helm-ai-kernel mcp receipts --json	List local MCP boundary records.
helm-ai-kernel mcp get --server-id <id> --json	Inspect one MCP server record.

No local command can turn an approver string or receipt-shaped value into an executable MCP approval. Servers remain quarantined until a credential verifier is wired to the governing approval authority.

Boundary Inspection

helm-ai-kernel boundary status --json
helm-ai-kernel boundary records --verdict ESCALATE --json
helm-ai-kernel boundary verify --record-id <record-id> --json

Receipts

helm-ai-kernel receipts status --format json
helm-ai-kernel receipts list --format json
helm-ai-kernel receipts show <receipt-id> --format json
helm-ai-kernel receipts verify --receipt <receipt.v5.json> --trusted-public-key-file <expected-ed25519.pub>
helm-ai-kernel receipts export --evidence DIR --out DIR
helm-ai-kernel receipts tail --agent <agent-id>
helm-ai-kernel workstation verify-decision --receipt <receipt.json>
helm-ai-kernel workstation verify-decision --receipt <receipt.json> --trusted-public-key-file <expected-ed25519-public-key>
helm-ai-kernel verify receipt --receipt <receipt.v5.json> --trusted-public-key-file <expected-ed25519.pub>

status, list, and show are bounded inspect. tail streams SSE and is refused as a listener inside the operator TUI. verify and export are aliases of the existing routes.

ALLOW, DENY, and ESCALATE records include a reason code. DENY and ESCALATE do not dispatch in enforce mode.

Workstation verification exits successfully only when receipt integrity and the signer trust anchor both verify. A signature that validates against the key embedded in a receipt is not, by itself, proof of an expected signer.

verify receipt is Foundation/offline verify for a Kernel receipt.v5 evaluate file. Exit 0 only when integrity and the caller-supplied --trusted-public-key-file both hold. It is not AI OS live, not helm-ai-kernel#859, not --allow-self-attested, and not workstation verify-decision. Hop fixtures are DENY / no permit.

Doctor

helm-ai-kernel doctor --format json
helm-ai-kernel diag --format json   # alias

Doctor reports PASS / WARN / FAIL checks and a healthy boolean. Suggestions point at inspect-first setup commands (setup status, setup repair … --dry-run); they do not recommend --yes. Exit 0 means no WARN/FAIL; exit 1 means WARN only; exit 2 means one or more FAIL.

OpenAI-Compatible Proxy

helm-ai-kernel proxy \
  --upstream https://api.openai.com/v1 \
  --port 9090 \
  --receipts-dir ./helm-receipts

Point an OpenAI-compatible client at http://127.0.0.1:9090/v1.

Format Contract

Operator-data commands accept --format text|json (legacy --json stays as an alias). Unknown formats exit 2. Collision verbs (verify, import, skills) keep domain --format meanings. Listeners and tui are exempt from emitting a JSON operator document. Details: CLI I/O Convention.

Help

helm-ai-kernel help
helm-ai-kernel help --all
helm-ai-kernel mcp --help
helm-ai-kernel verify --help