We checked 49 payment MCP servers: 37 can move money, 4 document a spending budget

Pink Agentic AI Payments is the only server in our 49-row census of payment MCP servers that documents all four core agent-spending controls (budget, per-payment limit, payee allowlist, and human approval) at once: 1 of 49. We publish this as PinkWallet, the company behind Pink, and Pink is scored from its public docs by the same criteria as every other row; it currently runs as a public sandbox with test money only. With that out of the way, here's what we found.

Why we counted this

If an AI agent can pay for things, something has to decide how much it can spend, who it can pay, and what happens when it tries to go over. MCP (Model Context Protocol) servers are becoming the default way agents get tool access, including payment tools, so we went looking for every payment-adjacent MCP server we could find and asked two questions of each one: can it actually move money, and if so, does it document any spending control.

We ended up with 49 servers: official vendor implementations (Stripe, PayPal, Square, Coinbase, Marqeta, and others) plus community and unofficial wrappers (x402 micropayment tools, Solana payment servers, third-party API wrappers). Every cell in the dataset carries a public evidence URL. Full methodology, findings and the raw CSV are on GitHub: github.com/Pink-Agentic-Payments/payment-mcp-census.

The headline numbers

• 37 of 49 (76%) servers can move money: create a payment, transfer, payout, charge, or issue a card. The other 12 (24%) are read-only, reporting, codegen, or discovery tools, even though several belong to payment companies (Mercury, Dwolla, Plaid, Lithic, Brex, and Circle's own codegen server are all classified as not moving money, by their own documentation).
• Of the 37 that move money, only 4 (11%) document a configurable spending budget: Marqeta, Gatewards, MCPay, and Pink.
• Only 5 (14%) document a per-payment limit: Coinbase AgentKit/CDP, Crossmint, MCPay, x402search-mcp, and Pink.
• Only 3 (8%) document a payee allowlist or block mechanism: Coinbase AgentKit/CDP, Marqeta, and Pink.
• Only 6 (16%) document a human-approval step: Stripe, Crossmint, Ramp, GoCardless, SpendNod, and Pink.
• Only 1 (3%) documents a spend-velocity rate limit distinct from a dollar budget: Marqeta. This is the least-covered control in the whole census.
• Only 8 (22%) document a dedicated sandbox or test mode.

The table

Here are the 11 money-moving servers that document at least one spending control, generated directly from the CSV:

"Not documented" means we didn't find a public mention of that control, not that the server definitely lacks it. The full 49-row table, with an evidence URL per cell, is in the GitHub repo.

Where Pink lands

Pink is the only row in the census that is a "yes" on all four core governance controls at once: budget, per-payment limit, payee allowlist, and human approval. Every other server documents at most two of those four. That is a real finding, not a coincidence of us grading our own homework: we used the same six-dimension rubric for Pink's row that we used for everyone else's, sourced only from Pink's public developer docs, and the row says plainly that Pink currently runs as a sandbox with test money, not production funds.

Pink Agentic AI Payments is an MCP server that checks per-agent budgets, payee rules and human-approval holds before a payment executes; you can try it in the public sandbox with test money: agentic-sandbox.pinkwallet.com.

The biggest gap: nobody rate-limits by time

Every control we counted caps how much an agent can spend, either per payment or in aggregate. Almost nobody caps how fast. Only Marqeta documents a velocity limit distinct from a dollar budget. That means an agent with a $10,000 monthly budget and no per-payment cap could, in most of these systems, spend the entire budget in one burst of calls in the first minute of the month. A budget that resets monthly says nothing about pacing. This looks like the most underbuilt control in the whole landscape, and a good place for the next vendor who wants to differentiate.

Five dead repos still listed as live

Five of the 49 servers in our census point to a GitHub repo that returned HTTP 404 when we checked on 2026-10-08, even though registries and mirror sites still list them as active: Fin-MCP, mcp-ap2, agent-pay, nullpath-labs/mcp-client, and agentshield-mcp. The GitHub accounts behind them still exist, so these look like deleted or renamed repos rather than banned accounts. If you're picking an MCP server off a registry listing, check that the repo actually loads before you build on it. Registry listings in this fast-moving space go stale quickly.

We also ran a direct query against the official MCP Registry API (not an aggregator mirror) for "payment" and "pay" and found 341 distinct server names across 210 publisher namespaces. The single largest namespace, app.wishpool, accounts for 81 of those names, one templated server per country wrapping a local payment rail. That raw list is in the repo too, as a cross-check, not folded into the scored 49.

How to correct a row

If you maintain one of these servers and we got something wrong, open an issue on the repo using the "Correct a row" template: name the server, the column, the correct value, and a public source. Corrections with a public source get merged within 7 days and logged in the changelog. Vendors fixing their own rows is the point of publishing this as a dataset instead of a one-off blog post.

Repo: github.com/Pink-Agentic-Payments/payment-mcp-census