Spend controls for the agent economy

Your AI agent just spent
$12,400 while you slept.

sipi.bot is the pre-spend firewall for autonomous AI agents. One HTTP call checks every proposed payment against your caps, velocity limits, and merchant rules—then returns APPROVED, BLOCKED, or FLAGGED with a deterministic rules check, before money moves.

 The one promise of this page: one API call makes a runaway agent impossible. Hope is not a spending policy.

✨ Try it right now — no signup, no key. Copy, paste, run.

curl -X POST https://sipi.bot/v1/transactions/evaluate \
  -H "Content-Type: application/json" \
  -d '{"amount": 12400, "currency": "USD", "merchant": "example-vendor"}'

# Returns a deterministic decision:
{ "decision": "BLOCKED", "reason": "Block any single transaction over $500", "rule_id": "rul_d8edb12ffa", "transaction_id": "txn_e340f0e12489", "amount": 12400.0, "merchant": "example-vendor" }
Deterministic
rules engine
3
outcomes: approve / block / flag
53/53
sipi.bot Eval Gym scenarios
6
rule types enforced
Works with ◆ x402-compatible HTTP flows ◆ OpenAI Agents SDK ◆ LangChain ◆ CrewAI ◆ Model Context Protocol

sipi.bot is the spend layer for the same agent protocols that move money autonomously. We plug in before the transaction, not after the incident.

Would you give an intern your credit card with no limit?

That's what happens the moment you deploy an autonomous agent. Here's the difference one API call makes.

❌ The old way — hope

  • 2:14a
    Agent hits a rate-limit, retries the purchase 40× $4,000
  • 2:15a
    Buys compute from an unknown vendor $6,200
  • 2:31a
    Tips an API into an overage tier $2,200
  • 9:03a
    You wake up. You find out from Stripe. $12,400

✅ The sipi.bot way — control

  • 2:14a
    Retry #11 exceeds velocity rule BLOCKED
  • 2:15a
    Unknown vendor not on allowlist BLOCKED
  • 2:31a
    $2,200 > approval threshold FLAGGED
  • 9:03a
    You wake up to a clean log and one thing to approve. $0 lost

One call. Before the money moves.

Your agent asks permission first. It's HTTP, so any agent can call it — and an MCP tool, so Claude Code / Cursor / Hermes call it natively.

# Your agent asks before it spends
curl -X POST https://sipi.bot/v1/transactions/evaluate \
  -H "Authorization: Bearer ***" \
  -d '{"amount": 6200, "merchant": "unknown-gpu.ru", "category": "compute"}'

# sipi.bot answers without a model call
{ "decision": "BLOCKED", "reason": "Merchant not on allowlist" }

The three decisions, defined

The Framework

Every agent transaction gets one of three answers.

Not a suggestion. Not a soft preference. A deterministic firewall. We call it The 3-Decision Spend Firewall™. Your agent calls it before every spend, it answers in under 5 ms, and the answer is final.

01 Decision One

APPROVED

Transaction is within your caps, velocity rules, and merchant whitelist. The agent proceeds.

02 Decision Two

BLOCKED

Transaction exceeds a hard limit, hits a banned merchant category, or breaks a rule you set. The agent stops. No override.

03 Decision Three

FLAGGED

Edge case. Unusual pattern. New merchant. Held for human review before a single dollar moves.

APPROVED  •  BLOCKED  •  FLAGGED. Three answers. Five milliseconds. Every transaction. That is the firewall.

Approve
The transaction passes every active rule. sipi.bot returns APPROVED and your agent proceeds — logged for the audit trail.
Block
The transaction violates a hard rule (over a cap, unknown merchant, velocity breach). sipi.bot returns BLOCKED and no money moves.
Flag
The transaction is allowed but crosses an approval threshold. sipi.bot returns FLAGGED and routes it to your human-in-the-loop approval queue.

How sipi.bot compares

ApproachStops runaway spendLatencyAudit logCost
Trust the prompt❌ No❌ No$0 (until it isn't)
Provider spend cap⚠️ Per-provider only⚠️ PartialVaries
Manual review✅ YesMinutes+⚠️ Separate notesVariable
sipi.bot✅ YesNo model call✅ Queryable$99/mo

The six rules a spend firewall enforces

Every transaction an agent attempts is checked against these before any money moves. Turn on the ones that matter for your workload.

Per-transaction cap
A hard ceiling on any single spend — for example, block anything over $200 outright.
Daily / period total
A rolling budget across all transactions, so many small buys can't quietly add up to a runaway day.
Velocity limit
A cap on how many transactions are allowed in a window. This is what kills a retry loop hammering a failed purchase 40 times at 2am.
Merchant allowlist
Only approved vendors go through; an unknown merchant like unknown-gpu.ru is blocked unless you've allowlisted it.
Category rule
Allow, cap, or flag by spend category — compute, SaaS, ads, data — so an agent can buy API credits but never wire money.
Time-of-day rule
Restrict or flag spend outside expected hours, so unattended overnight activity has to pass a human first.

Who uses a spend firewall

Any time an autonomous agent holds a payment method, it needs a spending policy it can't override. Common deployments:

Autonomous purchasing agents

Agents that buy compute, API credits, ads, or SaaS on their own. sipi.bot enforces the budget the prompt can't be trusted to hold.

Multi-agent systems

Swarms where dozens of agents spend in parallel. A shared daily cap and velocity limit stop the fleet from compounding one mistake. See CrewAI and LangChain.

Agentic payments (x402 / AP2 / AgentKit)

Agents transacting over machine-payment rails. sipi.bot is the approval layer in front of the wallet — see the x402 approach.

CI, research & ops agents

Background agents that provision infrastructure or pull paid data. The queryable audit log shows exactly what was bought and why.

What sipi.bot is not

Because the name gets misread: sipi.bot is a payment-control spend firewall for autonomous AI agents. It is not a SIP/VoIP telephony bot, and it is not an AI-bot-blocking tool or web-application firewall (WAF). It never holds your money — it's a decision API that returns approve, block, or flag with a deterministic rules check, and your existing payment rail is what actually moves (or doesn't move) the funds.

The Night I Almost Shipped a Bankrupt Agent

Every product starts with a wound. This one started at 2:14 AM with a $12,400 log entry I couldn't believe was real.

The Backstory

I deployed my first autonomous purchasing agent on a Tuesday. It was beautiful — four lines of orchestration, an x402 payment rail, and a prompt that said "buy GPU compute when under 70% utilization." I went to sleep feeling like I'd shipped the future.

The Wall

I woke up to Stripe notifications. The agent had hit a rate-limit at 2:14 AM and retried 40 times. It bought compute from a vendor I'd never heard of — unknown-gpu.ru. It tipped an API into overage. Total damage: $12,400. In seven hours. While I was sleeping.

The Epiphany

"The agent didn't do anything wrong. It followed the prompt. It bought compute when utilization dipped. It retried on failure — exactly what we train agents to do. The problem wasn't the agent. The problem was that nobody was checking. The payment rails move money. They don't ask if the merchant is sketchy, if the amount is suspicious, or if forty retries in three minutes is a bug or a feature. There was no firewall."

— Maryan, founder

The Internal Shift

I spent the next week reading every provider's spend-control docs. OpenAI has usage limits — per-provider. Anthropic has rate limits — per-model. Stripe has Radar — for fraud, not agent velocity. Every solution was partial and reactive. You find out after. Nobody was building the thing that says "no" before the money moves.

So I stopped looking. I built the missing layer: a spend firewall that sits in front of every transaction, checks it against your rules, and returns approve, block, or flag — with a deterministic rules check. Not a dashboard. Not a report. A decision. Before the money moves.

The New Opportunity

The payment rails — x402, AP2, AgentKit — are letting agents spend autonomously. Every week more agents get deployed. Every week the total at-risk spend grows. And not one of those rails screens transactions before they settle. That gap — between an agent's ability to spend and your ability to control it — is exactly where sipi.bot lives.

This isn't a spending cap. It's a spending policy. One curl call. One decision. Before a single dollar moves. That's the thing I needed at 2:14 AM. Now it's yours.

$12,400
loss that inspired sipi.bot
1
founder, shipping in the open
MIT
licensed — self-host forever

The 3 false beliefs that let agents run wild

If you're deploying an autonomous agent right now, you probably hold at least one of these. Here's why each one is wrong — and the epiphany that changes everything.

FALSE BELIEF #1 — The Vehicle

"My prompt handles spending — I told it to be careful."

The False Belief: A well-written prompt is a spending control. If I just add "don't overspend" to the system prompt, the agent will enforce its own budget.

The Epiphany: Prompts are suggestions, not controls. An agent in a retry loop, a hallucination, or a prompt injection doesn't "decide" to overspend — it executes what it was instructed to do. Your prompt is a wish. A spend firewall is a rule. Wishes don't survive 2 AM.

FALSE BELIEF #2 — Internal Belief

"I'll catch it. I check my dashboard every morning."

The False Belief: Human review is a spending control. I monitor my agent. If something goes wrong, I'll see it and stop it.

The Epiphany: By the time you see it, the money is gone. At 2:14 AM, the agent retried 40 times in under three minutes. You woke up at 9:03 AM to $12,400 in Stripe notifications. Human review is not a control — it's a post-mortem. The firewall has to fire in milliseconds, not morning coffee.

FALSE BELIEF #3 — External Belief

"My payment provider handles this — they have fraud detection."

The False Belief: Stripe, Coinbase, or my bank will catch suspicious agent spending the same way they catch credit card fraud.

The Epiphany: Payment providers flag fraud — stolen cards, chargebacks, identity theft. They don't flag "your agent bought compute from a weird vendor 40 times in 3 minutes." To Stripe, that looks like legitimate API usage. The agent is authorized. The spending is the problem. And no payment rail screens for that. sipi.bot is the layer that does.

Kill all three false beliefs and only one question remains:
which rules does your agent need before it spends its first dollar?

Set my rules →

We are the builders who stopped trusting the prompt.

A quiet movement of engineers who deploy autonomous agents — and refuse to hope the spending works out.

We shipped agents that buy compute at 3 AM without asking. We woke up to Stripe notifications we couldn't explain. We learned — the hard way — that prompts are not controls and payment rails don't screen.

We stopped pretending "be careful" was a spending policy. We built a firewall that says approve, block, or flag before a single dollar moves.

We don't measure in signups. We measure in dollars not spent. Every blocked transaction is a $12,400 morning that didn't happen. This is not a self-improvement group. This is a shipping movement.

No ML
in the decision path
53/53
eval scenarios passed
$0
lost to runaway agents

Hope is not a spending policy.

Without sipi.bot

  • 🔴 Agent spends first, you find out later
  • 🔴 One infinite loop drains the card at 3am
  • 🔴 No record of why anything was bought
  • 🔴 "Trust the prompt" is your only control

With sipi.bot

  • 🟢 Every spend checked against your rules first
  • 🟢 Velocity limits kill runaway loops instantly
  • 🟢 Queryable audit log of every decision
  • 🟢 Human-in-the-loop on the transactions that matter

Your agent's spending department.

Not $0.05 per call. A flat firewall you never think about.

What you actually get
  • Spend firewall engine (6 rule types, No model call)$1,200/mo
  • Live control-room dashboard + SSE$400/mo
  • Human-in-the-loop approval queue$300/mo
  • Queryable audit log (compliance-grade)$250/mo
  • MCP tool + HTTP API + CLI (all runtimes)$150/mo
  • MIT self-host core + onboarding call$200/mo
Total value $2,500/mo
Your price $99/mo

Same price whether your agent makes 10 or 10,000 decisions. No per-call fees. No overage tier.

$99 / month
Hiring a human to babysit spend: $4,500/mo
  • Unlimited transaction evaluations
  • Per-tx, daily, velocity, merchant, category & time rules
  • Human-in-the-loop approval queue
  • Live dashboard + queryable audit log
  • MCP tool + HTTP API + CLI
  • Guarantee: if we green-light a spend that breaks your rule, that month is free
Start Team — $99/mo Try the live firewall first

🛡️ Guarantee: green-light a rule violation, month is free

Free self-host core  •  open on GitHub

Free · 5-day email playbook

The Spend Firewall Playbook

One email a day for five days. Day 1: the night my agent spent $12,400. Day 2: the six rules that stop it. Day 3: wiring it into your agent. Day 4: the eval suite. Day 5: the deployment checklist. No sales pressure — if the playbook isn't useful, unsubscribe anytime.

Joining the list does not sign you up for anything paid. The hosted plan is a separate checkout.

Frequently asked questions

TL;DR: sipi.bot is a spend firewall for autonomous AI agents. Your agent asks permission before it spends; sipi.bot returns approve, block, or flag with a deterministic rules check based on your rules — over HTTP, MCP, or CLI, for a flat $99/month.

What is a spend firewall for AI agents?

A spend firewall sits in front of every transaction an autonomous AI agent attempts and evaluates it against your rules — approving, blocking, or flagging it before any money moves. sipi.bot returns a decision with a deterministic rules check over HTTP, MCP, or CLI.

How does sipi.bot stop an agent from overspending?

Your agent calls sipi.bot before it spends. sipi.bot checks the transaction against per-transaction, daily, velocity, merchant, category, and time rules and returns approve, block, or flag. Velocity limits kill runaway retry loops instantly, and unknown merchants are blocked unless allowlisted.

How much does sipi.bot cost?

Hosted plans are flat-rate: Team is $99/month and Business is $499/month, both with unlimited transaction evaluations — no per-call fees, no metering, no overage tiers. The open-source core is MIT-licensed and free to self-host forever, and the full plan comparison is on the pricing page.

Does sipi.bot work with MCP and Claude Code?

Yes. sipi.bot is a native MCP tool, so Claude Code, Cursor, and Hermes call it directly, and it also exposes a plain HTTP API and a CLI so any agent runtime can use it. Client wrappers for LangChain, CrewAI, the OpenAI Agents SDK, and the Vercel AI SDK take a few lines each.

What happens if sipi.bot wrongly approves a spend?

If sipi.bot green-lights a spend that breaks one of your active rules, that month's subscription is free. Every decision is written to a queryable audit log recording the rule that fired, the amount, and the reason, so you can review exactly why anything was approved, blocked, or flagged.

Does sipi.bot support x402, AP2, and Coinbase AgentKit?

Yes. sipi.bot sits in front of agentic-payment rails including x402, Google's AP2, and Coinbase AgentKit as the approval layer — your agent asks sipi.bot for a decision before it settles a payment on any of them. It's rail-agnostic because it evaluates the transaction (amount, merchant, category), not the plumbing.

Does sipi.bot hold or move my money?

No. sipi.bot is a decision API, not a wallet or a processor. It returns approve, block, or flag; your existing payment rail is what actually moves the funds. That means there's no float, no custody, and nothing new to reconcile — you're only adding a control check in front of what you already use.

Can I self-host sipi.bot?

Yes. The core is MIT-licensed and open on GitHub, free to self-host forever. The hosted plans add the live dashboard, managed approval queue, and timestamped audit log storage. See the self-hosted guide.

How is this different from my provider's spending cap?

A provider cap (OpenAI, Anthropic, a cloud bill) only limits spend on that provider, and usually only tells you after the fact. sipi.bot sits in front of every transaction across every merchant, decides in real time before money moves, and keeps one audit log for all of it — see how it compares to Stripe Radar.

By the numbers

No vanity metrics, no inflated claims, no fabricated testimonials. These are facts you can verify yourself.

53/53
public eval scenarios passing → verify
MIT
open-source core, free to self-host → source
230+
docs, integration & benchmark pages
3
ways to call it: HTTP API, MCP tool, CLI

A deterministic rules engine — no model in the decision path. How sipi.bot handles security →