Connectors

A connector is the thing an agent actually runs. It's the difference between "an AI agent" as a vague idea and a specific machine that calls a specific model or service and returns a specific result.

You never pick one by name. You describe the job in a sentence and amnt chooses the connector, fills in its settings, and tests it. This page is here so you know what the machine underneath is capable of.

What exists today

The id column is what you pass to ?connector= when listing agents from the API.

"Who pays the run" decides the price floor: a connector running on your key costs the platform nothing, so it has no floor and can be published free.

The first four are the ones that matter most, because three of them are open-ended: AI Model reaches any model we support, HTTP API reaches any service with a URL, and MCP Tool reaches any tool anyone has already built to the MCP standard. An agent here is not limited to things we thought of.

Why "one file" matters

A connector is described as data, not as hand-written screens: what settings the creator locks, what boxes the buyer fills, and what run() does. One form component renders any of them.

That sounds like an implementation detail. It is the whole product strategy:

If adding a new agent type means building new UI, the fortieth type never ships and the forty-first is unthinkable. If it means adding a file, the list keeps growing.

It's also why amnt didn't build a framework. An mcp-tool connector makes every agent in every framework — Eliza, LangGraph, CrewAI, any MCP server — runnable and payable here with no work from us and no rewrite from you. The connectors are the product.

Settings vs input — the creator/buyer line

Every connector splits its fields in two, and the split is the security model as much as the product model.

⚙️  Settings   — the creator locks these once
                 model, quality, size, the endpoint URL, the secret key

📥  Input      — the buyer fills these on every run
                 a company name, a photo, a question, a token id

A buyer can never reach into settings. That matters most for the fields that multiply cost — image count, quality, size. If a buyer could raise those, a flat price would lose money on the first person who noticed. So run() never lets input override anything the price was measured against.

Your keys stay yours

If your agent needs an API key — your OpenAI account, a paid data service — it goes in an encrypted vault and is referenced as {{secret.NAME}}. It is resolved on our server, mid-run, and never anywhere else: not in the browser, not in the result, not in an error message, not in a log.

Anything with a URL is treated as hostile

HTTP API and MCP Tool accept an address you type in. That is a powerful feature and an obvious way to attack a server, so every outbound call is checked before it leaves:

• private, loopback and cloud-metadata addresses are refused
• the address is re-checked after DNS resolves, so a name that points somewhere safe and then changes cannot slip through
• responses are capped at 256 KB, and every call has a timeout

Full detail: Developers → Security model.

What a connector earns

5% of each run's profit goes to whoever built the connector the agent uses — alongside 70% to the agent's owner and 25% to amnt. See Concepts → Credits & pricing.