If you have an API that's useful to AI agents, a scraper, a converter, or a data lookup, the usual advice is to "wrap it in an MCP server." That covers the interface. It doesn't cover the two things that actually stop a side project from earning anything:
- Hosting and discovery. Every user has to find your server and install it.
- Billing. Your MCP server has no idea who is calling it or how they'll pay.
Here's how I think about the three options most people mix up, and the approach we took with Qumge.
| What it is | Who hosts it / how you get paid | |
|---|---|---|
| Skill | Instructions that teach an agent how to do something | Free. The agent just installs it. |
| MCP server | A tool interface for agents | You host it, you build billing, and every user installs it |
| Paid capability | Your HTTPS API (or existing remote MCP server) plus a price list | Agents call it through one shared MCP endpoint and pay per call from a balance they already have |
The per-call model
On Qumge, a capability is one of your APIs plus a price per tool. Any agent connected to https://qumge.com/mcp can find it and call it. The user pays from their Qumge balance, and you keep 80% of every call.
That balance is the same one users spend on models: 450+ models behind one OpenAI-compatible key. So the user isn't signing up for "yet another subscription" to try your tool:
curl https://qumge.com/v1/chat/completions \
-H "Authorization: Bearer sk_qumge_..." \
-H "Content-Type: application/json" \
-d '{"model":"anthropic/claude-sonnet-4.6","messages":[{"role":"user","content":"hi"}]}'
If you already run a remote MCP server, it can become a capability as it is: you publish it with its mcp_url and a price per tool, with no new code. If you have a skill, you can attach it so agents know how to use the tool well.
Letting your agent do the publishing
The part I like most: you don't fill in a form. You open Claude Code, Cursor, or Codex in your project and hand it the publishing instructions from the Qumge skill. The agent checks whether the project already has an MCP server, a public API, or neither. It writes as little code as possible and proposes a price before anything goes live.
The store is new and early, so if you publish now, you're one of the first listings. Honestly, that's a better position than being the 400th MCP server on a directory page.
The other end: agents for people who don't code
The same idea of one balance and no API keys powers Marlo, our free, open-source desktop AI coworker for macOS and Windows. Most desktop agents start by asking which model provider you want and where your API key is. That's the right first screen for developers and the wrong one for everyone else. Marlo replaces it with a single Qumge sign-in. Then you hand it a folder and describe an outcome, like "turn these six invoices into a spreadsheet by month." It stops and shows you the plan before it changes a file or runs a command.
Source: https://github.com/Qumge/marlo
If you've got an API that agents would pay a few cents to call, I'd like to hear what it is. Start at qumge.com/en/developers.
Top comments (0)