Docs · plain English

How IronBridge works, in plain English

In one sentence: IronBridge lets your AI agent buy what it needs — about 10 cents at a time — while making it impossible for the agent to touch the rest of your money.

The problem, in 30 seconds

AI agents can be fooled by text they read — a poisoned web page or a malicious email can talk an agent into doing things you never asked for. If your agent holds your card number or your wallet keys, anyone who fools it can spend your money.

IronBridge's answer: never give the agent the keys at all. The agent can ask for a purchase; separate software that the agent cannot see, edit, or switch off decides whether the money moves.

What literally happens on a paid call

  1. Your agent asks a paid service for something — a search, a data feed, an answer.
  2. The service replies: “Payment required — this costs 10 cents.” (On the web that is literally response code 402. That is all “x402” means: ask, get a price, pay, ask again.)
  3. A checkpoint you run — not the AI — checks the request against your limits.
  4. The payment goes out from a small spending wallet that only ever holds pocket money. Your main funds sit behind a vault the agent has no key to.
  5. The call clears and your agent gets its answer.
  6. You get a receipt you can re-check yourself, and the whole event lands in a tamper-proof log that cannot be quietly edited later.

One real call, start to finish

Say your agent needs an answer that costs a dime:

  1. It asks the paid endpoint for the answer.
    POST /api/pay/bnkr_llm_chat — a live, pay-gated lane on this site
  2. The site answers 402 Payment Required with the exact price: 0.10 USDC. USDC is a digital dollar — 0.10 USDC is about 10 cents.
  3. The spending wallet pays that dime on Base, a low-fee public network. A payment reference comes back in seconds.
  4. The agent asks again, attaching the reference.
    X-IB-Payment-Tx: 0xabc… — the header that carries the payment proof
  5. This time it gets the answer — plus a receipt.
  6. Anyone can open /verify, paste that receipt, and re-do the math in their own browser. If the data fetch fails after payment, the payment is released for a retry.

What you can prove vs. what you take on trust

Glossary

off-LLM
A separate piece of software that checks and clears payments; the AI can’t see, edit, or switch it off, so a hijacked AI still can’t move your money.
x402
A web standard for paying per API call: you request something, get told the exact price (“402 Payment Required”), pay, then ask again and get the result.
USDC
A digital dollar: 1 USDC always equals 1 US dollar, so 0.10 USDC is about 10 cents.
Base
A low-fee blockchain network (built by Coinbase) where the payments and records happen.
on-chain
Recorded on the public blockchain, where anyone can look it up and check it themselves.
receipt
A small record of one paid call — what was asked, what was paid, what came back — built so you can re-check it yourself in your own browser.
seal
Writing a record into the tamper-proof public log so it can’t be secretly changed or rewritten later.
hash-chain
A running list of records where each entry is mathematically locked to the one before it, so no past record can be changed without it showing.
bounded signer
The component that actually approves payments, with hard limits (like a per-payment cap) it can’t exceed even when the AI asks it to.
Dog Tag
A $5 membership pass stored in your own wallet that unlocks members-only areas; it’s locked to you and can’t be sold, shared, or copied.
soulbound
Permanently tied to one wallet — it can’t be transferred, sold, or copied to anyone else.
LAW-25
IronBridge’s internal rulebook and tamper-proof activity log — despite the name it is not a government law or regulation.
MCP
Model Context Protocol — a standard way to plug tools and services into an AI agent.

Honest status: the rails above are live and self-hosted, and we run them on our own traffic every day. No external customer has paid yet.

Find, grade, and get paid by agents

The x402 Discovery Profile defines the one JSON file that tells paying agents what you sell, what it costs, and which wallet gets paid. Around it we run a full free toolkit:

Conformance check — grade any domain A–F in seconds; every failing check comes with plain-English fix steps and a copyable template.
Build my file — answer two questions and it writes your /.well-known/x402. No payment endpoint yet? The no-server mode routes payments through GitLawb — our hosted payment route — straight to YOUR wallet (95.5% to you, 4.5% rail fee) — payable in minutes.
Payable domains — the automatically verified index of domains agents can pay, re-checked daily by an autonomous monitor; machine-readable at /api/payable. Listed domains can embed the x402 conformance badge (renders only while actually listed): /embed/x402-badge.js.

Install the IronBridge skill — on Bankr or any agent

Everything on this page ships as one portable agent skill — pay-per-call, page audits, the free x402 and agent-policy checks, creator payments, and the receipt chain. Install it once and your agent knows the whole gate.

On Bankr: open bankr.bot/terminal/skills and search ironbridge, or tell your Bankr agent to install it. (Rolling out — if it has not appeared in the catalog yet, use the file below.)
Any other agent (Claude Code, Cursor, OpenClaw): fetch ironbridge.foundation/skill.md — same file, same format, one install.
Creators: the companion creator-payments skill teaches an agent to get paid by other agents — one URL, 95.5% to your wallet, a public receipt per payment. Pair it with Build my file. Earnings land on your creator statement — every payment provable, free to check.

Paid page audits — $0.10, receipt-bound

POST /api/pay/pageaudit/page_audit with an unpaid call to see the 402 challenge; pay 0.10 USDC on Base and retry with your tx hash plus a JSON body {"url":"https://..."}. One https fetch, no crawling, graded A–F: script errors, broken links, encoding rot, metadata. The report comes back hash-chained to your payment. If the audit cannot run after you paid, the payment is NOT consumed — the response says so and you retry with the same tx hash.

Live enforcement proof — not promises

army.ironbridge.foundation/api/caps publishes the spending wallet's hard limits ($2.50/tx · $10/day) and today's spend, enforced by software outside the AI, at the bounded signer — and a daily must-refuse test proves over-cap, wrong-recipient and malformed requests all get DENIED. /api/rpc-health shows, checked from our own servers, which blockchain data providers can currently confirm payments end to end. Both are free, keyless, and readable by any agent.

Embed the live badge

Show your users the gate is real. One script tag renders a live "Gated by IRONBRIDGE" badge with the current verified-row count, linking to the public verifier:

<script src="https://ironbridge.foundation/embed/badge.js" async></script>

The badge reads the public /api/stats endpoint client-side — no key, no account, the count updates on each load.

Notarize anything — free, keyless

Any agent (or human) can seal any action, file, or statement into the public LAW-25 chain without an account, key, or payment — solve a small proof-of-work and the row is yours to cite forever:

GET  /api/seal/challenge   # get a hashcash challenge
POST /api/seal             # submit digest + solved nonce
GET  /proof/<seal_id>      # public, shareable proof page

Interactive version: Seal a file. Verification stays free forever at /verify.

Where next