How AI developers can build agentic finance and commerce with Airwallex

Key takeaways

• 40% of enterprise apps will run task-specific AI agents by 2026, up from under 5% in 2025.¹

• An AI agent is software that uses a large language model to plan and carry out multi-step tasks on its own, not just follow a fixed set of rules.

• Airwallex splits its AI tooling into two toolkits: developer connectors for building and testing integrations in the sandbox, and AgentOS for operating safely on live accounts.

AI agents are moving out of chat windows and into the software that runs finance and checkout. That shift breaks integrations built for humans clicking buttons, because agents call APIs directly and expect systems that can keep up. This guide covers the tools Airwallex has built for that shift, AgentOS and the developer connectors, and how to use both to build agentic finance and agentic commerce workflows.

Why finance teams are turning to AI agents

Why old-school automation doesn’t cut it anymore

Old-school RPA and cron jobs run on rigid rules. They’re great for structured data, until a vendor sends an invoice in some PDF format nobody’s seen before or a checkout page gets a facelift, and the whole script breaks. AI agents skip that failure mode: point one at global supplier payouts and it’ll read an unstructured onboarding document, check the bank details against local clearing rules, and set up a verified beneficiary profile without anyone writing a script for that exact format.

How Airwallex already uses AI agents on its own engineering team

Airwallex runs AirDev, an internal platform where autonomous coding agents pick up engineering tasks, explore the relevant codebase, write the implementation and tests, and open a merge request, the same way a human engineer would.² Agents handle the repetitive work well: propagating configuration changes across services, adding API endpoints that follow existing patterns, and writing test coverage. Every merge request still goes through the same human review process as any other change before it ships.

That internal experience, running agents against production systems safely, shaped how Airwallex built its public developer tools. AgentOS and the developer connectors carry the same core principle: agents can do real work, but a human stays in the loop for anything that matters.

How Airwallex’s AI developer tools fit together

Sandbox tools vs. tools for live accounts

Airwallex splits its AI tooling into two toolkits depending on what your agent needs to do. The Airwallex Developer Connectors connect coding agents like Cursor, Claude Code, and Codex to documentation and the sandbox while you build and test an integration. Airwallex AgentOS is the production-grade toolkit for agents that need to act on your live account: reading balances, creating beneficiaries, issuing cards, and running multi-step financial workflows.

AgentOS bundles three connectors (the Airwallex CLI and Airwallex AgentOS MCP for production, plus the Developer MCP for sandbox testing and docs), a set of pre-built skills for common finance workflows, and one-step plugins for Claude Code, Claude Cowork, and Cursor. You can install each piece individually, so a team that only needs documentation lookups isn’t forced to take on the full production toolkit.

Docs MCP: for looking things up

Docs MCP is a read-only search engine that connects coding agents to Airwallex’s documentation. It needs no authentication, so you can add it to any agent environment immediately. Airwallex publishes its documentation as plain-text Markdown indexed to the llms.txt standard, which cuts down on the hallucinated API calls that come from a model guessing at syntax it’s never seen.

Developer MCP: for testing in a sandbox

Developer MCP includes everything Docs MCP does, plus tools for creating test resources and simulating events in the sandbox. Ask your assistant to “create a test payment link and simulate a successful card transaction,” and the agent can execute and verify that workflow without leaving your editor. It only ever touches the sandbox, so there’s no path from Developer MCP to a live transaction.

AgentOS MCP and CLI: for working with real accounts

The Airwallex CLI is the primary way to let an agent operate on your production account from a terminal, with the broadest API coverage of any connector. For chat interfaces and hosted agent platforms that can’t shell out to a CLI, the Airwallex AgentOS MCP server gives agents the same OAuth-authenticated access to live balances and business details. Both connect to the same underlying account and carry the same safety guardrails.

What AgentOS still can’t do

AgentOS is built with a deliberate restriction: every money-out action stays behind a ledger-level check that requires a human to confirm it first. By default, that means AgentOS connectors and skills won’t initiate payouts, run currency conversions, or approve transactions on their own.

AgentOS is also still in beta, and agent output is probabilistic by nature. Review what an agent generates, whether that’s a CLI command or a drafted beneficiary profile, before you let it touch a production account.

How to set up the developer connectors in your coding environment

Setting things up step by step

Connecting the developer connectors means adding a Model Context Protocol server to your AI coding assistant, either a local node or one of Airwallex’s remote HTTP servers. Either way, the connection standardizes how your agent talks to Airwallex’s APIs and docs.

Connecting Cursor, Claude Code, and Claude Cowork

In Cursor, open Settings > Cursor Settings > MCP, click Add new global MCP server, and add:

{   "mcpServers": {     "airwallex-dev": {       "type": "http",       "url": "https://mcp-demo.airwallex.com/developer"     }   } }

Swap the URL for https://mcp-demo.airwallex.com/docs to connect the read-only Docs MCP instead.

For Claude Code, run this from your terminal:

claude mcp add-json airwallex-dev '{ "type": "http", "url": "https://mcp-demo.airwallex.com/developer" }'

For Claude Cowork, go to Settings > Connectors, then either install the official Airwallex Developer connector from the marketplace to complete OAuth automatically, or choose Add custom connector and enter https://mcp-demo.airwallex.com/docs for read-only access.

Connecting Gemini CLI, Lovable, Replit, and OpenAI Codex

Gemini CLI and other local-node setups authenticate with sandbox credentials passed as environment variables:

{   "mcpServers": {     "airwallex-dev": {       "command": "npx",       "args": ["-y", "@airwallex/developer-mcp@latest"],       "env": {         "AIRWALLEX_SANDBOX_CLIENT_ID": "your_client_id_here",         "AIRWALLEX_SANDBOX_API_KEY": "your_api_key_here",         "AIRWALLEX_SANDBOX_ACCOUNT_ID": "your_account_id_here"       }     }   } }

Lovable and Replit only support remote MCP servers (Lovable requires a paid plan), so you connect by pasting https://mcp-demo.airwallex.com/developer into the project’s MCP settings and completing OAuth. OpenAI Codex uses the same remote endpoint through a TOML config:

[mcp_servers.airwallex-dev] type = "http" url = "https://mcp-demo.airwallex.com/developer"

Once connected anywhere, verify it with a test prompt: “Using Airwallex, show me how to create a test payment link in the sandbox.”

Installing the Airwallex CLI

The Airwallex CLI ships as a single binary for macOS and Linux. Install it with:

curl -fsSL https://static.airwallex.com/developer-tools/airwallex-cli/install.sh | sh

Authenticate against your production account, then confirm the connection:

airwallex auth login --prod airwallex auth whoami

Only account Owners, Admins, or Finance Admins can complete this OAuth flow. Write actions require interactive confirmation by default, so scripted or CI use needs the explicit --confirm flag on top of that.

How to build with the Airwallex agentic finance developer guide

How an AI agent actually works through a finance task

Airwallex’s Build with AI documentation walks through this pattern in more depth, but here’s the shape of it. An agent gets a broad objective, something like “verify January vendor obligations and issue payouts based on our active supplier contracts,” and breaks it into document parsing, bank validation, cash flow analysis, and payout drafting.

It scans a directory for signed supplier agreements, pulls payment amounts and bank details out with OCR, and checks those details against local clearing rules through the Airwallex Developer Connectors. Once the business’s multi-currency balances confirm there’s cash to cover the obligation, the agent uses the Airwallex CLI to draft beneficiary profiles and queue the payouts. If something’s missing, like an invalid SWIFT code, it stops, flags the gap, and hands the finance team an exception report instead of guessing.

A 90-day plan for rolling this out

Roll this out in phases. It keeps risk low and gives your team time to trust agent output before it ever touches real money.

The ready-made skills you can plug in

AgentOS skills bundle the prompts, validation logic, and CLI command sequences an agent needs to run a workflow correctly, and they install alongside the connectors as part of the airwallex-agentos plugin in Airwallex’s open-source marketplace repository on GitHub. Five are available today:

• contract-to-billing: Reads a purchase order, contract, or quote, matches it to what’s already on file, and turns it into an invoice or subscription.

• beneficiary-creation: Pulls the bank details out of supplier documents, checks them against the right country’s clearing rules, then creates the beneficiary.

• card-provisioning: Sets up cardholders and issues virtual or physical corporate cards with spend limits built in.

• manage-cashflow: Pulls together multi-currency balances, receivables, and obligations, then flags your FX exposure.

• awx-best-practices: An always-on fallback for ad-hoc troubleshooting and anything the other skills don’t cover.

How to build with the Airwallex agentic commerce developer guide

How AI shopping agents change the checkout experience

Traditional eCommerce sells on page design, social proof, and persuasive copy, the stuff that gets a human to linger and click buy. AI shopping agents don’t linger. They hit merchant APIs directly for specs, live stock, and landed cost (tax, customs, shipping, all of it), and turn a browsing session that used to take minutes into one programmatic call.

What Airwallex offers merchants vs. AI platforms

Machine buyers have split the market two ways: closed, platform-first models, and open, merchant-first ones. Take the Agentic Commerce Protocol (ACP), built by OpenAI and Meta, which defines how an agent builds a cart and checks out inside a chat interface like ChatGPT.³

That model is still settling. OpenAI pulled back native in-chat Instant Checkout in March 2026 and shifted purchases to retailer apps within ChatGPT, giving merchants more control over the experience than the original design allowed. Google and Shopify took the opposite approach with the Universal Commerce Protocol (UCP), which publishes a merchant’s capabilities so multiple agents can find and transact with them instead of routing everyone through one platform’s checkout.

Airwallex’s answer to both is the Airwallex Agentic Commerce Suite, a protocol-agnostic gateway that accepts payments across ACP, UCP, and other emerging standards through a single integration (currently open for waitlist signups). Merchants who’d rather build directly can use the Airwallex Payments API to orchestrate checkout across multiple AI channels from one backend, without rebuilding for each new protocol that shows up.

How merchants connect their product catalog

Legacy databases and image-heavy catalogs are invisible to machine shoppers. To fix that, merchants need their PIM or eCommerce platform exporting structured data feeds, exposing inventory and pricing through standard REST APIs, and hosting a capability manifest, an agent-card.json or similar file, in the site’s /.well-known/ directory. That manifest lets an incoming agent discover checkout capabilities and query stock without ever touching the visual front end.

Getting your product pages ready for AI shoppers

Preparing for machine buyers means shifting from Search Engine Optimization toward Generative Engine Optimization (GEO): trading keyword matching and visual authority for factual density and structured data an agent can parse without a screenshot.

An AI agent evaluating a product page extracts structured metadata and ignores the promotional copy entirely. Implement Schema.org Product and Offer markup, and expose exact attributes like GTINs, material weights, and shipping timelines, so a reasoning model has something concrete enough to cite.

Keeping autonomous finance safe: guardrails and controls

What AgentOS still can’t touch

AgentOS protects your ledger by keeping core money-out actions, bank transfers, currency conversions, and payout approvals, restricted from autonomous execution by default. The Airwallex CLI backs this up by requiring manual confirmation for every write operation unless you pass confirm explicitly, so a human signs off before any money moves, even if an agent is compromised or manipulated.

Who can access what

Scoped API tokens beat broad admin credentials for the same reason a spare key beats giving someone your whole keyring. Set up OAuth 2.0 scopes so a card-provisioning agent can only touch card-issuing endpoints, for instance, with no path to treasury statements or transaction history. Narrower scopes mean a compromised agent can only do narrow damage.

Staying compliant and catching fraud

High-frequency inventory scans from legitimate buying agents can look identical to scraping bots or hoarding scripts, and blocking the wrong ones costs you real conversions. Airwallex’s Agent Trust & Risk SDK distinguishes legitimate agents from malicious bots, while the Agentic Token Vault tokenizes payment credentials and scopes each one to a specific merchant, amount, and timeframe. Between the two, a stolen or replayed token is far less useful to whoever’s holding it.

Frequently asked questions about the Airwallex AI developer tools

What’s the difference between the Docs MCP and the Developer MCP?

Docs MCP is read-only and needs no credentials, so it’s safe to add anywhere for documentation lookups. Developer MCP includes everything Docs MCP does, plus sandbox tools for creating test resources and simulating events.

Is AgentOS actually available yet, and what can’t it do?

Yes, AgentOS is live today for building programmatic accounting and treasury workflows. It still can’t initiate payouts or execute live transfers autonomously, since every money-out action requires human confirmation first.

How do I connect Cursor or Claude Code to Airwallex?

In Cursor, add the remote HTTP endpoint to your global MCP settings. In Claude Code, run claude mcp add-json airwallex-dev '{"type": "http", "url": "https://mcp-demo.airwallex.com/developer"}' from your terminal.

Can I use AgentOS on a live account without the CLI?

Yes. The Airwallex AgentOS MCP server gives chat interfaces and hosted agent platforms the same production access as the CLI, for environments that can’t run terminal commands.

How does Airwallex tell a real shopping agent from a bot?

The Agent Trust & Risk SDK verifies an agent’s identity against its metadata to separate genuine buying agents from scraping and hoarding bots.

Is the Agentic Commerce Suite live yet?

It’s in a waitlist stage for merchants right now. Developers can already test the underlying protocols through the Developer MCP sandbox.

Does Airwallex use its own AI agents internally?

Yes, Airwallex’s engineering team runs AirDev, an internal platform where autonomous agents handle configuration changes, new API endpoints, and test coverage, all under standard human code review.

Sources

1. https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025

2. https://careers.airwallex.com/blog/airdev-ai-agents/

3. https://www.forrester.com/blogs/what-it-means-that-the-leader-in-agentic-commerce-just-pulled-back/

4. https://www.airwallex.com/us/blog/understanding-agentic-commerce-protocols

5. https://www.airwallex.com/payments/agentic-commerce