Prevent double charges + runaway spend

Your AI will eventually trigger an expensive incident. OnceOnly prevents it.

Q: Do you have a Python SDK?

Yes — Python SDK + REST API. See docs at https://docs.onceonly.tech/sdks/python/.

Q: How do billing and refunds work?

Subscriptions renew monthly and are processed securely by Stripe. You can cancel anytime. We offer a 14-day refund — see the Refund Policy page or email support@onceonly.tech.

OnceOnly sits between your agent and real side-effects (Stripe, emails, CRM, APIs). It blocks duplicates, caps spend, and enforces tool permissions.

Start free — prevent your first incident Read docs

No credit card. 2 minutes.

Decision API demo

5-minute integration

Python SDK

client.gov.upsert_policy({
  "agent_id": "billing-agent",
  "allowed_tools": ["stripe.charge"],
  "max_spend_usd_per_day": 50.0
})

key = f"charge:order:{order_id}"
if client.check_lock(key=key, ttl=3600).duplicate: return

client.ai.run_tool(agent_id="billing-agent", tool="stripe.charge", args={...}, spend_usd=3.50)

Duplicate charges across retries
Forbidden tools (allow/deny)
Runaway spend (hard caps)

Why we built OnceOnly

A short founder-style note and the real scenarios this prevents in production.

Why we built this

We built OnceOnly after seeing AI agents double-charge users and loop paid APIs in production.

Every AI system that can act in the real world will eventually do something twice.

Real scenarios we stop

Stripe double charges / duplicate refunds
API retry storms burning budget overnight
Agents mutating or deleting the wrong record

Put OnceOnly right before Stripe/email/CRM writes. One call. Instant safety.

Fast setup (minutes) Exactly-once side-effects Budgets + allowlists Audit logs

Start free. No credit card. 2 minutes.

What goes wrong without OnceOnly

AI agents retry, race, and loop. When they touch money or external systems, the failure mode is expensive.

Without OnceOnly

WITHOUT

Double charges / double refunds
Runaway spend overnight
Tool loops → API bans / outages
Refunds + chargebacks + angry users (trust damage)
No audit trail when things break

Outcome: refunds + chargebacks + lost trust

With OnceOnly

WITH ONCEONLY

Exactly-once side-effects
Hard budgets + rate limits
Tool allowlists/denylists
Audit logs: executed / blocked / deduped + why

Outcome: dedup + caps + audit trail

One AI retry storm can cost more than a year of OnceOnly.

Real production failure

A concrete breakdown: trigger, impact, and how it gets prevented.

Incident: Duplicate refunds

Real incident

Trigger

AI support agent retried refund logic (retries + races).

Impact

34 duplicate refunds → $8,200 lost + 2 days support cleanup.

Prevention

OnceOnly: idempotency key + budget cap + audit logs.

Time to integrate: 5 minutes

What to implement: action key + check + cap

Where teams use OnceOnly

Common production use-cases where AI agents and automations create real risk.

Teams use OnceOnly to protect

Payment agents
AI support bots
Automation platforms
Multi-worker AI jobs

Before vs after

Before
LLM → Stripe → Stripe → Stripe → INCIDENT

After
LLM → OnceOnly → Stripe (1x)

How it works

OnceOnly is a decision API. Before your agent performs a real action, it asks OnceOnly. Only one run gets permission.

Pick an action key

Example: charge:order:123

Use stable IDs you already have (order_id, event_id, user_id).
Reuse the same key on retries.

Ask OnceOnly

Allowed or duplicate — one decision before the action.

Allowed → run once.
Duplicate/locked → skip safely or poll.

Execute side-effect once

Stripe, email, CRM, API writes happen exactly once.

Do the irreversible call once.
Failures are safe to retry with the same key.

AI agent / automation

OnceOnly decides

Stripe · Tools · Emails · APIs

Allowed → run once. Duplicate/locked → skip or poll safely.

Built for production AI bot teams

Everything you need to ship faster without burning budget or breaking systems.

Exactly-once actions

Guarantee a side-effect happens once, even with retries, race conditions, and webhook replays.

AI leases

One agent acquires and executes; everyone else polls. Prevents double execution under load.

Tool governance

Allow/deny tools per agent, add pricing rules, and hard caps so agents can’t “invent” dangerous actions.

Budgets & rate limits

Cap actions/hour and spend/day to stop runaway loops and tool spam.

Audit logs

Get a forensic trail of decisions: executed, blocked, deduped, and why.

Kill switch

Instantly disable a rogue agent (Agency plan) during an incident.

Python SDK: ship in minutes

If you can call HTTP, you can use OnceOnly — but the SDK makes it fast.

Install

pip install onceonly-sdk

Then use check_lock(), AI leases, and governance APIs.

Python SDK docs View on PyPI

Governed tool call

from onceonly import OnceOnly

client = OnceOnly(api_key="onceonly_sk_...")

client.gov.upsert_policy({
  "agent_id": "support-bot",
  "allowed_tools": ["send_email"],
  "max_actions_per_hour": 120,
  "max_spend_usd_per_day": 25.0
})

res = client.ai.run_tool(
  agent_id="support-bot",
  tool="send_email",
  args={"to": "user@example.com"},
  spend_usd=0.001
)

ROI in one incident

Examples (not promises) of what a single AI incident can cost when it hits money, paid tools, and production. Examples from typical production failures (loops/retries).

Duplicate payments

When an agent processes payments, duplicates become direct loss:

20 duplicates \times $200 = $4,000

Plus refunds and support time. Even 2 duplicates × $200 = $400 + support time.

Runaway tool loop

An overnight loop calling paid tools / LLMs can rack up:

$500-$5,000 unexpected spend

Incident response

One on-call incident:

3 engineers \times 4 hours

Plus trust damage and time spent on postmortems.

Incident Type	Typical Loss	With OnceOnly
Duplicate payments	$5K–$50K	$0 (exactly-once)
API loop	$500–$40K	Hard spend cap
Tool misuse	Data loss / bans	Tool allowlists
Incident debugging	6–12 eng hours	Full audit log

Pro ($149) is insurance. Agency ($799) is incident control. If it prevents one incident, it paid for itself.

Pricing

Pick a plan based on your risk. If your agent touches money or writes to external systems, choose Pro.

Pro includes governance (policies + tools registry) and audit logs. Agency adds higher limits and a kill switch for incident response.

Free

Prototypes and testing

$0 / month

Idempotency checks: 1,000 / month
AI leases: 3,000 / month
TTL default: 60s
TTL max: 1h
Governance + audit logs: —
Tools registry: —
Kill switch: —

Get API Key

Starter

Low-risk bots

Shipping your first production bot

$29 / month

For first production bots. Stops duplicates.

Idempotency checks: 20,000 / month
AI leases: 100,000 / month
TTL default: 1h
TTL max: 24h
Governance + audit logs: —
Tools registry: —
Kill switch: —

Choose Starter

Pro

💳 Bots touching money

Best for most AI teams

$149 / month

If your agent touches money or writes to external systems, Pro is the default.

Idempotency checks: 200,000 / month
AI leases: 1,000,000 / month
TTL default: 6h
TTL max: 7d
Governance + audit logs: Included
Tools registry: Up to 10 tools
Kill switch: —

Choose Pro

Agency

🚨 Mission-critical AI

High-scale production + kill switch

$799 / month

For teams where one incident > $10k. Kill switch + high limits.

Idempotency checks: 2,000,000 / month
AI leases: 10,000,000 / month
TTL default: 24h
TTL max: 30d
Governance + audit logs: Included
Tools registry: Up to 500 tools
Kill switch: Included

Choose Agency

Metric	Free	Starter	Pro	Agency
Idempotency checks / month	1,000	20,000	200,000	2,000,000
AI lease requests / month	3,000	100,000	1,000,000	10,000,000
Default TTL	60s	1h	6h	24h
Max TTL	1h	24h	7d	30d
Tools registry	—	—	Up to 10 tools	Up to 500 tools
Agent governance (policies + observability)	—	—	Included	Included
Kill switch	—	—	—	Included

OnceOnly is cheaper than one AI incident.

Stripe billing: Renews monthly • Cancel anytime • Secure payments via Stripe. We offer a 14-day refund — see Refund Policy or email support@onceonly.tech. Taxes may apply where required. Charges may appear as OnceOnly.

FAQ

Quick answers for teams shipping AI agents in production.

How much can one AI incident cost?

It depends on what your agent can touch. Incidents commonly include duplicate payments, refund/chargeback overhead, and runaway tool spend — ranging from hundreds to tens of thousands of dollars.

Can one incident cost more than a year of OnceOnly?

Yes. A single duplicate payment batch or overnight retry loop can exceed the yearly cost — plus the hidden cost of support time and trust damage.

What types of failures does this stop?

Duplicate side-effects (charges, refunds, emails), retry storms, tool loops, and unsafe tool usage. OnceOnly enforces exactly-once actions, budgets, and tool allow/deny rules.

Where do I put OnceOnly in my stack?

Right before the irreversible call: Stripe, email send, CRM write, or any API mutation. Your agent asks OnceOnly first; only one run is allowed to execute.

How fast can we integrate?

Minutes. Add one call before the side-effect and reuse the same action key on retries.

Do you store our payment data or business payloads?

No. OnceOnly does not store payment data or business payloads — only action keys and decision metadata needed for dedupe, governance, and audit logs.

Do you have a Python SDK?

Yes — Python SDK + REST API. Start here: Python SDK docs.

How do billing and refunds work?

Renews monthly • Cancel anytime • Secure payments via Stripe. We offer a 14-day refund — see Refund Policy or email support@onceonly.tech. Charges may appear as OnceOnly.

Is it safe to put in the critical path?

Yes. OnceOnly is built to make low-latency decisions so you can put it directly in front of irreversible calls.

How does it work (technical)?

OnceOnly is a decision API in front of side-effects: it blocks duplicates, enforces budgets/tool permissions, and records audit logs. See the docs for full details and the API reference: docs.onceonly.tech.