We are building payment infrastructure specifically designed for AI agents in the EU.
The problem with giving an agent access to standard payment APIs (like Stripe) is that LLMs are probabilistic—if they hallucinate a JSON payload or get stuck in a loop, they can drain a bank account and also obviously human trust.
To solve this, we built an architecture specifically for "Agentic" workflows. It forces strict schema validation and defaults to a Human-in-the-Loop flow for now (the agent drafts the payment, but generates a consent_url for a human to approve).
Where we are at:
We do have the production rails connected and moving real money. However, because the security implications of agentic payments are so high, we are heavily limiting production seats.
Instead, we just opened up a fully mocked Sandbox. I want developers to try locally before we ever let their agents touch real euros.
You can drop the built-in MCP server/toolkit into your workflow in 30 seconds:
pip install whire (works with python 3.11+)
(Use the public mock key: whire_test_key)
I would love your feedback on the API ergonomics, the tool-calling logic, and any blind spots we might be missing.
Happy to answer any questions!
Hey HN, I'm the founder of Whire.
We are building payment infrastructure specifically designed for AI agents in the EU.
The problem with giving an agent access to standard payment APIs (like Stripe) is that LLMs are probabilistic—if they hallucinate a JSON payload or get stuck in a loop, they can drain a bank account and also obviously human trust.
To solve this, we built an architecture specifically for "Agentic" workflows. It forces strict schema validation and defaults to a Human-in-the-Loop flow for now (the agent drafts the payment, but generates a consent_url for a human to approve).
Where we are at: We do have the production rails connected and moving real money. However, because the security implications of agentic payments are so high, we are heavily limiting production seats.
Instead, we just opened up a fully mocked Sandbox. I want developers to try locally before we ever let their agents touch real euros.
You can drop the built-in MCP server/toolkit into your workflow in 30 seconds:
pip install whire (works with python 3.11+)
(Use the public mock key: whire_test_key)
I would love your feedback on the API ergonomics, the tool-calling logic, and any blind spots we might be missing. Happy to answer any questions!