CDMX
Agentic finance. Payments in 80+ currencies, and an AI CFO that runs the back office with a person on every send.
01 — The idea
Blaze started as a better way to send money across borders. It settles over stablecoin rails and pays out on local ones like ACH, SEPA, SPEI and PIX, in more than 80 currencies.
Then we pointed AI at the boring half of running a business. The AI CFO reads a company's bank, accounting and invoice data and does the work: pays bills, drafts invoices, creates payment links, pays contractors. It lives in the Blaze Business dashboard, and it plugs into ChatGPT and Claude as a remote MCP connector.
I co-founded Blaze and run engineering. I wrote the first commit of the current monorepo and I'm still its top committer.
02 — The walkthrough
01 · Welcome
Blaze is for people whose money lives in more than one country. Paid in dollars, rent in pesos, a client in Europe. The first screen says it plainly.
02 · Home
Your balance shows in the currency you live in, with USDC underneath. Everything starts from four shortcuts: receive, manage your accounts, make a payment link, send to a contact.
03 · Get paid
A US account for ACH and wire, a Mexican CLABE and a European IBAN for SEPA. Share the details and get paid like a local, wherever the money comes from.
04 · Or skip the banks
Receive USDC on Stellar, Ethereum or Solana. Copy an address or show a QR code, and it lands in the same balance.
05 · Send
Money settles over stablecoin rails and pays out on local ones like ACH, SEPA, SPEI and PIX. Every step shows up in your transaction history.
06 · Insights
Monthly spend against your budget, a heads-up when a category jumps, and spending broken down by country.
07 · Widget
A home screen widget with live exchange rates, so you know what a dollar is worth in pesos before you open the app.
03 — What it does
A dashboard for bills, invoices, customers, products, coupons and accounting, with owner statements and insights on top.
86 tools across balances, bills, receivables, forecasting and runway, reconciliation, payroll, accounting reports, FX and payouts. Ask it what you can afford this month and it answers from your own books.
The same agent ships as a remote MCP server with OAuth scopes. 20 tools read, 4 propose, 2 execute, and every proposal hands back a link for a person to confirm. There's a CLI too, @blaze-money/cli.
Websites as a service for the same customers who get paid through Blaze: restaurants, barbers, studios, auto spas.
04 — In the hand
05 — Under the hood
Confirm methods are never registered as agent tools. The model proposes a payment, a person confirms it in the UI, and only then does the server execute it. Every payment the agent starts is tagged as agent-initiated.
Owners can let the agent pay small bills. The limit is the smaller of 1% of available balance or $500, computed per payment on the server, along with a 24-hour cool-off for new payees, a duplicate check and the bill policy engine. Fail any gate and it falls back to asking.
The harness runs the real agent against seeded staging data and scores tool routing, arguments and numeric grounding: any dollar figure that can't be traced back to tool data gets flagged. CI runs only the scenarios a pull request touches, comments the result, and drafts new scenarios for tools nothing covers yet. The MCP suite hard-fails if the model ever says a proposal was paid.
The Vercel AI SDK routes across Bedrock, Vertex, Azure and OpenAI, with a direct Claude driver alongside. The business always comes from server context, never from model input, and MCP output is projected and redacted per entity.
Built with
06 — What I took from it
The model is the easy part of an AI CFO.Most of the work is deciding exactly where it has to stop and ask a person.
Contact
I like talking to founders early — before the deck, before the team. Tell me what you're working on.
Write to me