Let your agents pay.
Keep the wall standing.
Parapet is developer-first, consumer-ready infrastructure that lets AI agents make payments on your behalf — verified, policy-bound, and durable. We believe agentic payments can be more secure than traditional rails. This is the standard we're building to.
AI has become hugely ubiquitous in our lives — from asking it for dinner recipes and health questions, to researching the latest products. The explosion of OpenClaw and Claude Cowork has only increased our reliance on AI to not just help us with work, but to do work.
But this new age of working and discovering products is stunted the moment it meets a traditional issue — one we treat as trivial in physical life: making payments. It's incredibly easy to pay now, with digital wallets on phones and even watches. The agentic experience doesn't share this. Despite there being strong desire, with 31% of UK consumers aged 18–34 willing to let agents manage their purchases entirely autonomously.
Parapet solves this. It's developer-first, consumer-ready infrastructure that empowers your agents to make payments on your behalf. Naturally, change in banking and payments is needed — but Parapet pushes the boundaries of what's doable now, while setting the standard for what's needed in the future.
& TRUST
Traditional payments run on assumed trust. Agentic payments don't have to.
This thesis centres on two words: verification and trust. I'd argue traditional payments have little of either.
- The person paying with a card is assumed to be its owner, or at least authorised to use it.
- The merchant is assumed to be correct — the EFTPOS machine, the website's payment portal.
There are checks, particularly online — SSL certificates on merchant sites, biometric push notifications from banks. But a lot of that trust, and therefore verification, comes from social cues and context. Does this store seem legitimate? Did the shopkeeper swap machines? Does this payment portal look familiar?
Parapet holds the belief that agentic payments can be more secure and more reliable than traditional payment rails. This is the standard we aim to achieve.
Let's look at some examples. A straightforward, relatively verifiable intent is a consumer using Parapet's MCP server to enable payments inside their favourite LLM — managing their finances with little extra work. Every payment request routed through the MCP server passes through Parapet's payment execution, which we cover further down.
From there, it's easy to imagine agentic commerce removing the human who triggers the payment entirely.
MCP server, inside your LLM
A consumer connects Parapet's MCP server to their assistant and manages day-to-day spending conversationally, with every request routed through Parapet's execution layer.
"Buy it when it's on sale"
A user gives an instruction to Hermes Agent or OpenClaw — the guidance is human, but the purchasing action itself is triggered by the agent or application layer.
Employee expense claims
Businesses provision Parapet to handle claims. With integrations to Xero, Sage and Quickbooks, claims reconcile in seconds to minutes — not days to weeks.
These examples rely on strong agent verification — answering the question: is this agent who it says it is? This will be an ever-evolving space. Today, Parapet achieves it with OAuth 2.1 and PKCE exchange, which lets agents be authenticated and access revoked at any time.
Once an agent is authenticated, next comes delegation mandates — this agent is authenticated, but do they have access to this wallet?
PARAPETS
Guardrails, enforced at transaction time — not after the fact.
A unique thing about agentic payments is the idea of policies, or guardrails. Unlike traditional rails, individual or organisational policy can be enforced at the moment of the transaction itself.
By the time a policy is reviewed, we already know a lot: we've confirmed the agent's identity, and confirmed the agent can access this wallet. What's left is confirming whether the agent is allowed to make this payment, from this wallet.
This is where spending limits and other policies come in. Policies in Parapet can be agent-specific, but also span across wallets and organisations.
A simple policy: cap spend at £100 a day. A more complex one: each employee gets a £100 monthly budget, but the organisation's overall spend can't exceed £10,000 a month.
Policies can go further still — down to the type of business or person you wish to pay. Only allow payment to approved suppliers, say, or businesses within the automotive parts industry.
Policies are designed for you to sleep easy at night.
INTENT
Identity and access aren't enough. Someone still has to ask: was this actually meant?
Policies aren't the only guardrail. The other missing piece is payment intent. For a human-initiated transaction, intent is relatively clear. For an agent-initiated one, it's much less clear.
Parapet providing the payment service shouldn't be the sole goalkeeper here — but Parapet should be a goalkeeper. The LLM developer, or the application layer more broadly, should help gather genuine payment intent. Parapet then independently verifies that intent, checking the conversation transcript for prompt injection or deception. This is what stops an agent buying an £80 product when the user's instruction was £50.
Guardrails are a joint effort: the application layer works to enforce them and avoid errors, for a positive user experience; Parapet works to detect errors, for safe payment rails. Every interaction, approved or denied, is logged and visible to users in the Parapet control plane.
If users connect accounting or ERP software to Parapet, contacts are synced as potential payees. That raises confidence that a payment lands with:
- Someone who passes the UK's Confirmation of Payee
- Bank details that match a user-owned third-party system, such as accounting software
Parapet also checks whether the user — or anyone else on the platform — has paid this payee before. These checks build a real network of information that increases the stability and safety of agentic payments for everyone involved. If there's any doubt, we reject the payment or send it for biometric approval on the Parapet control plane.
Where integrations exist, Parapet closes the loop on reconciliation. Every transaction syncs back to your accounting integrations, giving complete visibility and finishing the accounting workflow — not just the payment.
RELIABLE
EXECUTION
The worst thing Parapet could do is pay someone twice.
To prevent that, payments have to be durable and fault-tolerant. That's why all payment execution runs on DBOS — handling unexpected errors, seeking approval on the Parapet control plane, and managing execution and reconciliation end to end.
Payment intents are also deliberately short-lived. Look at Parapet's MCP server and you'll find two tools:
Runs every check and prepares the request — but doesn't move money. Returns a very short-lived intent_token.
Requires that exact intent_token to run. No token, no execution.
An LLM must call prepare_payment() first. To execute, that exact intent_token must be used — which eliminates tool hijacking, where a bad actor or a compromised LLM tries to modify a payment request after Parapet has already approved it.
TO CHANGE
Regulation assumes a human in the loop. Agentic payments don't have one.
Payment regulation naturally assumes a human is present in the process. With agentic payments, that human oversight doesn't exist — which means the default fallback of biometric authentication (SCA) falls flat.
Ultimately, payment regulation needs to evolve to recognise agents as first-class entities. Google's Agent Payments Protocol (AP2) is a thesis toward that change.
As stated throughout this piece, we believe agentic payments can be more secure than traditional rails — decreasing the reliance on SCA and strict payment regulation over time. Parapet builds on current regulation and helps shape what comes next: the foundation for consumers, businesses and builders to bring payments into their workflows and applications, without having to worry about compliance.
CRYPTO
We're taking a bet on fiat rails — while respecting the case for crypto.
We're not the only movers here. Cloudflare has recently introduced cloudflare.pay, and alongside Coinbase is attempting to revive the x402 protocol — originally introduced in 1999 to let APIs request payment. We agree that stablecoins, and cryptocurrency more broadly, are well suited to this problem.
There's a wider argument for crypto as the agent's native currency: it certainly has less friction with compliance and payment regulation. That said, Parapet is taking a bet that consumers and businesses will prefer to move money over fiat rails.
That's why Parapet exists — to enable reliable agentic payments over fiat rails.
Building the wall your agents can pay through.
We're onboarding a first group of developers and finance teams ahead of launch. Tell us what you're building, and we'll bring you in early.