veritera 0.4.0

Forge Verify SDK — Verify AI agent decisions with cryptographic attestation

Forge Verify SDK for Python

Verify every AI agent action before execution. Cryptographic attestation. Sub-15ms latency. Zero data access.

Install

pip install veritera

Complete Flow: Zero to Verified in 60 Seconds

from veritera import Forge

forge = Forge(api_key="vt_live_...")  # Get your key at forge.veritera.ai

# 1. Create a policy (do this once)
policy = forge.create_policy_sync(
    name="finance-controls",
    description="Block high-value transactions and dangerous operations",
    rules=[
        {"type": "action_whitelist", "params": {"allowed": ["payment.read", "payment.create", "balance.check"]}},
        {"type": "amount_limit", "params": {"max": 10000, "currency": "USD"}},
        {"type": "action_blacklist", "params": {"blocked": ["database.drop", "admin.override"]}},
    ],
)
print(f"Policy created: {policy.name} (ID: {policy.id})")

# 2. Verify an action
result = forge.verify_sync(
    action="payment.create",
    agent_id="finance-bot",
    params={"amount": 500, "currency": "USD", "recipient": "[e-mail]"},
    policy="finance-controls",
)

if result.verified:
    print(f"Approved — proof: {result.proof_id}")
    # Safe to execute the action
else:
    print(f"Blocked — reason: {result.reason}")
    # Do NOT execute

# 3. List your policies
for p in forge.list_policies_sync():
    print(f"  {p.name} (v{p.version}) — {len(p.rules)} rules")

# 4. Test a policy without executing
test = forge.test_policy_sync(
    policy_id=policy.id,
    action="database.drop",
    params={"table": "users"},
)
print(f"Test: {test.verdict}")  # "denied"

That's it. No GUI needed. No dashboard required. Everything from code.

Async Usage

import asyncio
from veritera import Forge

async def main():
    async with Forge(api_key="vt_live_...") as forge:
        policy = await forge.create_policy(
            name="email-controls",
            rules=[{"type": "rate_limit", "params": {"max_per_hour": 50}}],
        )

        result = await forge.verify_decision(
            agent_id="support-bot",
            action="send_email",
            params={"to": "[e-mail]", "subject": "Your refund"},
            policy="email-controls",
        )
        print("Approved" if result.verified else f"Blocked: {result.reason}")

asyncio.run(main())

Generate Policies from Natural Language

Don't want to write JSON rules? Describe what you want in plain English:

result = forge.generate_policy_sync(
    "Only allow my agent to read files, send emails (max 50 per hour), "
    "and check balances. Block all deletions and admin operations.",
    save=True,  # save it immediately
)
print(f"Created: {result['name']}")
print(f"Rules: {result['rules']}")

Use with Framework Integrations

The policy you create here works with all Forge framework packages:

# Create the policy once
forge.create_policy_sync("finance-controls", rules=[...])

# Then use it in any framework:
# OpenAI Agents SDK:  forge_protect(tools, policy="finance-controls")
# LangGraph:          ForgeVerifyMiddleware(policy="finance-controls")
# CrewAI:             ForgeVerifyTool(policy="finance-controls")
# LlamaIndex:         ForgeVerifyToolSpec(policy="finance-controls")

Package	Install
forge-openai	pip install forge-openai
langchain-forge	pip install langchain-forge
crewai-forge	pip install crewai-forge
llama-index-tools-forge	pip install llama-index-tools-forge

API Reference

Forge(api_key, **options)

Option	Type	Default	Description
api_key	str	required	Your API key (vt_live_... or vt_test_...)
base_url	str	https://forge.veritera.ai	API base URL
timeout	float	10.0	Request timeout (seconds)
max_retries	int	2	Retries on 5xx errors
fail_closed	bool	True	Return denied instead of raising on errors
debug	bool	False	Enable debug logging

Verification

Method	Description
verify_decision(agent_id, action, params, policy)	Verify an action (async)
verify_sync(action, agent_id, params, policy)	Verify an action (sync)
get_proof(proof_id)	Retrieve a verification proof
verify_proof_locally(attestation, payload, public_key)	Verify Ed25519 attestation offline

Policies

Method	Description
create_policy(name, rules, description)	Create a new policy (async)
create_policy_sync(name, rules, description)	Create a new policy (sync)
list_policies() / list_policies_sync()	List all active policies
get_policy(policy_id) / get_policy_sync(policy_id)	Get a policy by ID
update_policy(policy_id, name, rules, description) / update_policy_sync(...)	Update a policy
delete_policy(policy_id) / delete_policy_sync(policy_id)	Deactivate a policy
test_policy(policy_id, action, params) / test_policy_sync(...)	Test a policy without persisting
generate_policy(prompt, save) / generate_policy_sync(...)	Generate policy from natural language
get_policy_templates()	Get all available policy templates

Policy Rule Types

Type	Description	Params
action_whitelist	Only allow specific actions	{"allowed": ["action1", "action2"]}
action_blacklist	Block specific actions	{"blocked": ["action1", "action2"]}
amount_limit	Cap transaction amounts	{"max": 10000, "currency": "USD"}
rate_limit	Limit action frequency	{"max_per_hour": 50}
time_window	Restrict to business hours	{"start": "09:00", "end": "17:00", "timezone": "US/Eastern"}
require_confirmation	Flag for human approval	{"actions": ["payment.create"]}
recipient_constraint	Control who agent can contact	{"allowed_domains": ["@company.com"]}
resource_access	Restrict file/resource access	{"denied_resources": [".env*", "*.key"]}
custom	Custom constraint logic	{...}

Delegations

Method	Description
create_delegation(agent_id, allowed_actions, constraints, expires_in)	Create a scoped delegation

Account

Method	Description
get_usage(period)	Get billing usage statistics
health()	Check API health

Error Handling

from veritera import ForgeError, RateLimitError

try:
    result = await forge.verify_decision(...)
except RateLimitError as e:
    print(f"Rate limited — retry in {e.retry_after_ms}ms")
except ForgeError as e:
    print(f"Error: {e.code} ({e.status}): {e}")

Fail-Closed (Default)

When fail_closed=True, network/server errors return a denied result instead of raising. Your agent is blocked, not crashed. This is the safe default.

Circuit Breaker

After 5 consecutive failures, the SDK opens a circuit breaker for 30 seconds. After 30s, one request is allowed through (half-open). On success, the circuit closes.

Requirements

• Python >= 3.9
• Dependencies: httpx, cryptography

Get Your API Key

1. Go to veritera.ai
2. Sign up (free tier: 250 verifications)
3. Copy your API key from the dashboard
4. Set it: export VERITERA_API_KEY=vt_live_...

License

MIT — Forge by Veritera AI