AI Agents

Banking your agent can run.

Open accounts, issue cards, pay invoices, run payroll — through a Claude, Codex, or Cursor conversation. One install, one API key, one clear set of guardrails.

Install in Claude Code Read the API docs
brighty-agent
> agent: check balance for account EUR
✓ Balance: €12,450.00
> agent: send €500 to DE89370400440532013000
✓ SEPA transfer initiated. ETA: instant
> agent: exchange 1000 USD to EUR
✓ Exchanged $1,000.00 → €921.50 (rate: 0.9215)
> agent: pay for invoice ~/Downloads/service_invoice_120226.pdf from usdc
✓ Invoice paid: $2,340.00 from USDC account
> █

What it does

Open accounts

CURRENT for spend, SAVING for vault, in any currency Brighty supports.

Send payments

SEPA, SWIFT, ACH, on-chain. One transfer or a batch of thousands.

Pay invoices

From a PDF or a forwarded email. The agent extracts the fields and waits for your confirmation.

Run payroll

Drop a CSV, the agent stages every transfer, you review, then commit.

Issue cards

Virtual or physical. Set the spend policy, freeze on suspicion, unfreeze on confirmation.

Convert balances

Live FX between your own accounts at the real Brighty rate. Preview the quote before committing.

Manage team

Invite teammates, set roles (Owner / Admin / Payer / Viewer), remove members.

Read account details

IBAN, BIC, SWIFT, on-chain addresses — fetch what you need to share with a payer.

Send to a Brighty username

Move money between Brighty businesses by username. No IBAN, no fees, settles instantly.

How It Works

01

Install once

Two commands in Claude Code, or one MCP config block in any other client. The agent gains 24 banking tools and four skills that teach it when to use them.

02

Set your key

A Brighty API key from your Business Portal. Per operator, per machine. Never goes through the agent.

03

Ask in plain language

The agent drafts the call. You confirm before any money moves.

A real conversation

What it looks like when your agent uses Brighty. Tool names are real. The confirmation gate is not optional.

claude — brighty transcript
You
Show me the EUR account balance.
Agent
Calls brighty_list_accounts. 
EUR Current  €12,403.50
EUR Savings  €50,000.00
You
Pay this invoice. (attaches PDF)
Agent
Extracts the invoice fields. Shows you:
Recipient:  Acme Industries GmbH
IBAN:       DE89 3704 0044 0532 0130 00
Amount:     €1,234.56
Reference: INV-2026-04-001
Source:     EUR Current (€12,403.50)
Confirm?
You
Yes.
Agent
Calls brighty_create_payout, then brighty_create_external_transfer, then brighty_start_payout after you re-confirm the staged transfer.

Install in your client

Pick the path that matches what you're already using.

/plugin marketplace add razz-team/brighty-agent-toolkit /plugin install brighty@brighty-agent-toolkit
  1. 1. Run both commands in Claude Code or Claude Desktop.
  2. 2. Set BRIGHTY_API_KEY in the shell that starts Claude, then restart it.
  3. 3. Verify with /mcp — the brighty server should show as connected.
Full install guide → Source on GitHub → Get an API key →

Designed for trust

Operator-confirmed

Every money-moving call is gated by your client. The agent drafts the request with all resolved fields — IBANs, amounts, source accounts — visible verbatim.

You commit. The agent does not.

No silent transfers, by design

Key never travels

Your API key lives in your env or OS keychain. The local MCP server reads it, attaches it as an Authorization header, and never logs it.

No tool in the toolkit takes a credential as an argument.

Credential isolation, by design

Open and auditable

MCP server and skills are open source. Every tool call is visible in your client’s transcript.

Source on github.com/razz-team/brighty-agent-toolkit.

Inspect everything, hide nothing

What stays under your control

The operator decides every payment. The agent never moves money on its own.

Every money-moving call requires explicit confirmation from the operator before the request reaches Brighty. Your client gates the call; the toolkit makes the gate visible by surfacing every resolved field — IBANs, amounts, currencies, source accounts — verbatim, before the call.

The gated tools include brighty_start_payout, brighty_transfer_own, brighty_create_external_transfer, brighty_freeze_card, and brighty_terminate_account.

Your API key never passes through the agent.

It's read from your environment or your OS keychain by the local MCP server, attached as the Authorization header on the way to Brighty, and never logged. No tool in the toolkit accepts a credential as an argument — by design.

If you run unattended, leave auto-approve off for these.

We recommend keeping auto-approve off for any tool whose name starts with brighty_create_, brighty_start_, brighty_terminate_, or brighty_freeze_. If you turn auto-approve on for those, you accept the corresponding risk.

What we don’t do (yet)

Things you won’t find in this toolkit. Naming them is more useful than promising them.

3-DS approvals

The agent cannot approve a 3-DS challenge for you. You handle those in the Brighty app.

Card termination

Exposed in the Brighty app, not in the toolkit. Freeze is reversible; termination isn’t.

Hosted MCP

The server runs locally as a subprocess of your client. There is no Brighty-hosted MCP endpoint and we’re not planning one.

Hand your agent the keys.

Two commands in Claude Code, or one MCP block elsewhere. Open source, operator-confirmed, ready when you are.

Install in Claude Code Read the install guide
Source on GitHub → Open an issue → Get an API key →