Enforcement started
August 2.
Zabta evaluates every agent action against 14 jurisdictions and answers with a verdict — allow, deny, or escalate — with the article that decided it.
12 policies
34 policies
19 policies
“The data subject shall have the right not to be subject to a decision based solely on automated processing … which produces legal effects concerning him or her.” — GDPR Art. 22(1)
policy eu_art22_optout · layer 02 JURISDICTION · matched is_automated_decision · 231ms · routed to human review
Illustrative evaluation — the policy slug, citation, and verdict are real seed data; the agent is an example.
NOTICES — IN FORCE
2026-08-02
High-risk AI system obligations under the EU AI Act became enforceable.
Penalties reach €35M or 7% of worldwide turnover, whichever is higher — EU AI Act Art. 99. Many high-risk deployers also owe a Fundamental Rights Impact Assessment before deployment — EU AI Act Art. 27. If your agents make automated decisions, generate content, or touch personal data in the EU, these articles may now apply to them.
2026-06-30
The Colorado AI Act took effect for developers and deployers of high-risk systems.
Consequential decisions by AI now carry a duty of reasonable care — Colorado SB24-205. US state law is arriving jurisdiction by jurisdiction; Texas TRAIGA is already on the rack below.
NOT READY TO SIGN UP? READ FIRST.
What the EU AI Act requires of your agents →THE ENGINE — HOW A VERDICT IS MADE
Every action crosses three registered layers over one fixed grid. Each layer prints its policies onto the evaluation; the most restrictive verdict wins. Nothing moves — the action is what travels.
12 policies
Safety baselines that apply everywhere: kill switch, audit trail, PII guardian, human approval for irreversible actions, cost limits.
human_approval · escalates is_irreversible
34 policies · 14 packs
The law where you operate. Activate the packs that match your footprint; each policy cites the article it operationalizes.
eu_art22_optout · GDPR Art. 22
19 policies · 5 overlays
Industry rules on top: Financial Services, Healthcare, Employment, Education, Critical Infrastructure.
hipaa_minimum_necessary · HIPAA
Illustrative records — policy slugs, citations, and verdicts are real; agents and latencies are examples within measured bounds (<500ms cached).
CONNECT — TWO LINES OF CODE
AUTO MODE
import zabta
zabta.init("zbt_your_api_key")
# That's it — OpenAI, Anthropic, and LangChain calls are now governed.MANUAL MODE
from zabta import ZabtaClient
client = ZabtaClient(api_key="zbt_your_key")
result = client.evaluate("process_refund", "order_123", {
"has_pii": True,
"is_irreversible": True,
})
if result.allowed:
process_refund(order) # policy allows it
elif result.escalated:
queue_for_review(order) # a human decides
else:
log_denial(result.reason) # denied — the reason carries the citationTHE BROKER — CREDENTIALS ARE GOVERNED TOO
BETA 0.1.0b5 — WIRE PROTOCOL MAY CHANGEPolicies decide what agents may do. The Broker decides what they may hold. A local daemon vaults your API keys; agents check out short-lived, policy-evaluated leases at the moment of use — the secret never sits in an env file or a prompt.
VAULT
Secrets encrypted at rest on your machine. The master key lives in your OS keychain. The secret never leaves the box.
AES via Fernet · keychain-held key
CHECKOUT
The agent requests a credential; the request runs the same policy pipeline as any action — identity, policy, budget.
zabta.credential("stripe", ["charges:create"])
LEASE
Allowed requests get the secret with a 300-second lease record. Denials get the reason. Every path is audited.
ttl 300s · fingerprint-only persistence
THE LEDGER — EVERY DECISION, RECORDED AND CHECKABLE
| time | agent | action | verdict | policy | latency |
|---|---|---|---|---|---|
| 14:02:11 | refund-agent | refunds.approve | ESCALATE | eu_art22_optout | 231ms |
| 14:02:09 | support-bot | tickets.read | ALLOW | audit_trail | 94ms |
| 14:01:55 | marketing-bot | content.generate | DENY | eu_watermarking | 187ms |
| 14:01:41 | prospector-2 | email.send | ALLOW | tool_allowlist | 102ms |
| 14:01:38 | data-sync | records.export | ESCALATE | pii_guardian | 204ms |
| 14:01:22 | deploy-bot | deploy.production | ESCALATE | impact_assessment | 176ms |
Illustrative records — the format is the product's; policy slugs are real, agents are examples.
CONTINUOUS — RUNS ON EVERY PUSH
256 tests across four suites
backend 50 · broker 51 · sdk 35 · policy 120
PRODUCTION RUN — MAY 2026
82/82 policy regression against live production
14 jurisdictions · 5 overlays · sub-500ms cached
The proof is checkable, not claimed: the SDK is live on PyPI, every jurisdiction policy carries its article, and the audit trail your agents generate is the same format you see above.
THE OPERATIONS ROOM — WHERE A HUMAN WATCHES THE BOARD
Escalations land in an approval queue. Compliance scores, trust monitoring, and the audit trail live in one dashboard — this one.
THE RACK — 14 JURISDICTION PACKS, REAL CITATIONS
Activate the packs that match your footprint. When jurisdictions conflict, the strictest applicable policy wins — configurable per tenant.
PRICING — START FREE, NO CARD
KNOW
Free
$0 /mo
See what your agents are actually doing.
- Cloud policy evaluation
- Up to 2 jurisdiction packs
- 3 agents · 1 seat
- 7-day audit history
- Community support
- —Sectoral overlays
- —Broker credential governance
- —Compliance dashboard
GOVERN
Starter
$99 /mo
Full regulatory coverage, plus local credential governance.
- All 14 jurisdiction packs
- All 5 sectoral overlays
- Broker credential governance
- Compliance dashboard
- 25 agents · 3 seats
- 90-day audit history
- Email support
- —PDF compliance export
PROVE
Pro
$399 /mo
For a team that has to prove compliance to someone.
- Everything in Starter
- PDF compliance & evidence export
- Unlimited agents · 10 seats
- 1-year audit history
- Priority support
SHAPE
Design partner
Talk to us
Deploying agents now and want to shape the tool that governs them.
- Early access to what's next
- A direct founder line
- Pricing locked in as we grow
- A limited number of spots
No credit card required · free tier forever