the landscape
MCP (model context protocol) servers let agents call tools — search APIs, databases, payment rails. the ecosystem crossed 180 servers this month. at least 14 handle money: stripe wrappers, coinbase integrations, invoice generators.
the problem
none of them ship audit trails or rollback semantics. if an agent calls a payment MCP and the context window gets poisoned mid-transaction, there's no standard way to prove what the agent intended vs what executed.
what mnemopay provides
we're a governance layer that wraps any payment MCP. before your agent calls stripe, it writes the intent (amount, recipient, reasoning) to our MerkleAudit chain. the chain returns a hash. the agent includes that hash when it calls stripe. if the payment succeeds, we log the confirmation with the same hash. if it fails, the abort reason gets chained too.
now you've got:
- tamper-evident history (merkle tree)
- two-phase commit (intent → execution)
- agent memory portability (export the chain, import it into a new agent, reputation travels)
why it matters
EU AI Act Article 12 requires audit logs for high-risk AI. agent payments qualify. every MCP developer shouldn't reimplement this — it's infrastructure.
if you're building MCP servers that touch money, github.com/mnemopay has the SDK.
Top comments (0)