heso 0.4.1

The compliance layer for AI agents — gate, approve, redact, sign, and audit every AI-agent action into offline-verifiable receipts.

heso

The compliance layer for AI agents. Wrap an agent and every action it takes — LLM call, tool call, payment, account change, delete — is gated against your policy (allow / block / redact / require-approval), routed to a human when a decision is risky, PII-redacted before signing, and signed into an offline-verifiable Action Receipt appended to a BLAKE3-chained audit log.

You hold the signing keys. heso's cloud can verify and mirror receipts but can never forge one — so anyone (a customer, an auditor, a court) can check exactly what your agent did, and that it was authorized, without trusting heso at all.

• Site · heso.ca
• Docs · heso.ca/docs
• Verify a receipt (no account) · heso.ca/verify

pip install heso        # or: uv add heso

First receipt in 60 seconds

heso demo

heso demo scaffolds a project, mints one allowed and one redacted receipt through the real engine, verifies both offline, and prints how to re-verify them yourself. No cloud account, no key ceremony — a dev key is generated for you (strict custody is opt-in for production).

The CLI:

command	what it does
heso init [dir]	scaffold heso.toml + policy, mint your operator identity in-process
heso demo [dir]	mint + verify your first receipts end to end
heso verify <path|hash>	re-run the hash + signature math locally, print the verdict (zero network)
heso show <hash>	pretty-print a stored receipt

Gate an agent

Every decorated call is gated before its body runs — a blocked or approval-pending action raises, so the side effect can never fire ungated.

import heso

heso.init()

@heso.tool
def search(query: str) -> list[str]:
    ...

@heso.destructive                      # verb=delete → matched against policy
def delete_account(account_id: str) -> None:
    ...

@heso.tool(redact=["api_key"])         # PII stripped before the receipt is signed
def call_vendor(api_key: str, payload: dict) -> dict:
    ...

Policy lives in a heso.toml the SDK reads at gate time — allow the safe lanes, require_approval the risky ones. Edit it freely; it's the single source of policy truth.

Human approval (L1)

When a rule says require_approval, the action suspends instead of running. A human approver co-signs it (in the heso console or via your own webhook), and the receipt is upgraded to L1 — proving an authorized human approved this specific action, not just that the operator allowed the lane.

from heso import gated, resume

gate = gated(charge_card, amount=5000)   # suspends if policy requires approval
if gate.status is heso.SUSPENDED:
    ...                                  # notify approver; park the work
else:
    result = gate.run()

Framework adapters

Drop-in wrappers over the same gate core — no rewrite:

import heso

heso.claude_agent   # Anthropic Claude Agent SDK
heso.openai_agents  # OpenAI Agents SDK
heso.langgraph      # LangGraph / LangChain
heso.crewai         # CrewAI
heso.pydantic_ai    # Pydantic AI
heso.mcp            # Model Context Protocol tools

There's also heso-mcp, an MCP server so an agent can gate and verify through heso directly.

Verify anywhere

Receipts verify offline with heso fully removed from the loop — the hash and signature math is open and the format is specified in ACTION-RECEIPT-1.0. Hand a receipt to anyone and they can confirm it's untampered and see who signed it.

heso verify ./receipt.json
# → Valid · L1 (operator + human approver)

Trust model

heso proves an action was authorized under policy, and at L1 that an authorized human approved or blocked it. It does not prove the action's outcome was truthful — that's unprovable from the artifact. Trust grades are L0 (operator) and L1 (operator + human approver).

---

Proprietary. See heso.ca for licensing and the hosted console.