statis-ai

The trust layer for production AI agents. One decorator turns any function into a policy-gated, audit-logged tool call. No agent framework lock-in, no proxy in front of your model, no rewrites.

statis-ai is the Python SDK for Statis. Drop it in front of any side-effectful function your agent calls — Stripe refunds, internal APIs, database writes — and Statis evaluates a deterministic policy before the function executes, then writes a signed receipt of the decision.

from statis import gate

@gate(action_name="issue_refund")
def issue_refund(charge_id: str, recipient_email: str, cents: int) -> dict:
    return stripe.Refund.create(charge=charge_id, amount=cents)

The decorator does three things on every call:

1. Context In — Strips prompt injections, redacts PII, meters tokens before the model call. (Re-exported from statis-kit, our open-source pre-call hygiene library.)
2. Action Out — Sends the proposed call to the Statis policy engine. Approved → executes. Denied → raises ActionDeniedError with a receipt. Escalated → human approves in the console; agent resumes.
3. Receipt Through — Every decision is recorded in an Ed25519-signed, hash-chained ledger. Verifiable offline with the public key at /.well-known/aarm-pubkey.

Install

pip install statis-ai
# optional: receipt signature verification (offline)
pip install "statis-ai[verify]"

Set your API key:

export STATIS_API_KEY=st_...

(Get one for free during beta at statis.dev.)

Quick start — gate a function

from statis import gate, ActionDeniedError

@gate(
    action_name="issue_refund",
    entity=lambda: {"customer_id": "CUST-7781"},
    target_system="stripe",
    timeout_s=10,
)
def issue_refund(charge_id: str, recipient_email: str, cents: int) -> dict:
    """Issue a Stripe refund. Only runs after the policy engine approves."""
    return {"id": f"re_{charge_id[-6:]}", "amount": cents, "status": "succeeded"}


try:
    result = issue_refund(
        charge_id="ch_3PxOrder9921",
        recipient_email="alice@bigcustomer.com",
        cents=4999,
    )
    print("approved:", result)

except ActionDeniedError as e:
    print(f"denied by rule {e.receipt.rule_id}: {e}")
    # e.receipt has receipt_id, hash, signature — audit-ready, no extra wiring.

Policy as code

Rules are deterministic YAML, evaluated server-side per call. No LLM in the decision path:

- rule_id: refund_recipient_must_match_state
  action_type: issue_refund
  conditions:
    recipient_matches_customer_of_record: true
  decision: APPROVED
  priority: 100

Apply with the CLI:

statis policy apply policies.yaml

What you get out of the box

• @statis.gate decorator — sync + async, idempotent, with timeout + escalation
• StatisClient — lower-level access for custom flows
• process_context() — re-exported from statis-kit for one-line pre-call hygiene (Guard + Compressor + Meter)
• CLI — statis init, statis policy apply, statis receipts list, statis receipts verify
• Distributed lock + at-most-once execution semantics
• Console (web UI) for approvals, escalations, kill-switch, and receipt verification

Free during beta

The Developer Cloud (policy engine, ledger, console) is free while in beta — no Stripe at launch. Self-hosted Enterprise (SSO, SCIM, VPC, compliance exports) is contract-driven.

Links

• Homepage → statis.dev
• Source → github.com/statis-ai/statis-core
• OSS kit → statis-kit (the Context-In primitives — works offline, no API key required)
• TypeScript SDK → npm install statis-ai

License

MIT.