Why we shipped our docs as an MCP server

By Venly Finance | July 20, 2026

We exposed docs.venlyfinance.com as an MCP server. An AI agent can now compose a working Venly integration from live docs in about a minute. Here's why — and how to try it.

Half of API traffic now reportedly comes from AI agents rather than humans typing code. The Model Context Protocol — donated to the Linux Foundation's agentic AI effort in late 2025 — has passed 10,000 public servers. Banks have started shipping MCP servers. Cloudflare publishes an "Agent Readiness" score.

Here's the thing about payment APIs in that world: an AI agent that guesses your endpoints is dangerous. An AI agent that reads your live documentation is useful. The difference is whether you built for it.

What we did

docs.venlyfinance.com is exposed as an MCP server, and our docs publish an llms.txt index. Connect it to Claude, Cursor, or any MCP-aware tool, and the agent stops hallucinating our API — it queries the actual surface: the real POST /accounts, the real virtual-bank-account fields, the real status enums.

The practical effect: an agent can compose a production-shaped integration plan — endpoints, request bodies, sequence diagram, cited doc pages — in about a minute. We know because that's how one of our partner conversations started: three doc servers connected, one prompt, a working marketplace-payout-with-stablecoin-leg architecture before the coffee arrived.

Try it (the actual prompt)

Add our docs MCP to your agent and run:

You have the Venly Finance docs MCP connected (docs.venlyfinance.com). Design a marketplace seller payout with a stablecoin leg: onboard a seller (POST /accounts with inline party), issue an EUR virtual IBAN that auto-converts to USDC (POST /accounts/{accountId}/virtual-bank-accounts), and pay out via POST /accounts/{senderAccountId}/transfers/crypto. Produce a sequence diagram labelled with exact endpoints and cite the doc pages used.

Every endpoint in that prompt resolves against our live docs. Check us.

Why this matters more for payments than for most APIs

In most domains, an agent's hallucinated field name costs a failed request. In money movement it costs a mis-routed payout, a broken idempotency chain, a compliance gap. Payment APIs that want agent traffic — and it's coming whether they want it or not — need three things: a machine-readable source of truth (MCP/llms.txt), marketing pages whose code examples match the docs exactly, and unambiguous enums and idempotency semantics that an agent can't misread.

We rebuilt our own site to that standard this quarter: every snippet on venlyfinance.com is validated against the docs MCP. Not because it's tidy — because agents read marketing pages too, and an agent that learns your API from a stale landing-page example will write stale integrations.

FAQ

What is an MCP server? A standard interface that lets AI tools query external systems — in this case, our documentation — instead of relying on training data.

Is this a security risk? The docs MCP is read-only over public documentation. Credentials and API calls stay in your environment.

Which tools work with it? Any MCP-aware client: Claude, Cursor, and a growing list of IDEs and agent frameworks.

Does Venly charge for it? No. It's documentation.