What Is an Agent skill.md & Agent-Native API?

If you landed here after seeing something like "give your agent our skill.md" or "agent-native API" and thought what does that actually mean? — this is the plain-English definition, one term at a time. It is deliberately a glossary entry, not a tutorial. If you want the step-by-step version, read the procedural companion guide. Here we just define the vocabulary so the rest of it clicks.

What is a skill.md?

A skill.md is a single Markdown file, published at a public URL, that documents an API specifically so an AI agent can read it and act on it directly.

That last part is the whole point. Traditional API docs are written for a human developer: you read them, understand them, then write code that calls the API. A skill.md collapses that loop. The file is written to be ingested by the model itself — endpoints, authentication, request and response shapes, and copy-paste examples, all in plain text — so the agent can learn the interface and start making correct calls without a person translating the docs into code first.

You can see a real one at agents.modelvi.com/skill.md. Open it in a browser and it looks like readable documentation. Hand the same URL to an AI agent and it becomes an instruction set the agent can follow.

What is an "agent-native API"?

An agent-native API is an API designed on the assumption that the caller might be an autonomous agent, not only a human developer clicking through a dashboard.

In practice that shows up as a set of design choices that make the API predictable enough for a model to use reliably:

• One stable base URL — e.g. https://modelvi.com/api/partner/v1 — so the agent never has to guess where to send requests.


• Simple, consistent auth — a single bearer token (Authorization: Bearer mvk_...) on every call.


• A consistent response envelope — every response comes back in the same shape, { success, payload }, so the agent parses results the same way each time.


• A public probe endpoint — an unauthenticated GET /version the agent can hit to confirm the service is reachable before doing anything else.


• Explicit rate limits — a documented ceiling (here, 600 requests per minute per key) and a machine-readable Retry-After header on an HTTP 429, so an agent knows exactly how long to back off instead of retrying blindly.


• Machine-first identifiers — stable codes like ONLYFANS or FAN (Fansly) rather than display names that might change.

None of these are exotic. An agent-native API is mostly just a well-disciplined API — but the discipline is aimed at a reader that can't improvise around inconsistency the way a human can.

What is a "self-serve API key"?

A self-serve API key is a credential you create yourself, in a dashboard, without a sales call or manual onboarding.

In the ModelVI example the key looks like mvk__, minted from the "API Keys" area of the dashboard. Two properties matter for the definition:

• Tenant-scoped — the key only reaches your own models and accounts, never anyone else's.


• Deny-by-default — it grants nothing until access is explicitly attached, which is what makes handing a key to an agent reasonable rather than reckless.

"Self-serve" is the operational difference from a legacy integration: a human sets up the credential in minutes, and the same key is what an agent (or a partner) then uses to authenticate.

What are "agent-readable docs"?

Agent-readable docs is the umbrella term for documentation written to be consumed by a model rather than only by a person — plain Markdown, at a stable URL, with the endpoints, auth, and worked examples all present in the text. A skill.md is one concrete format of agent-readable docs. The distinction from ordinary docs isn't the content so much as the intent: nothing is hidden behind a login, a rendered widget, or a "contact us to get started" wall that a model can't click through.

How the pieces fit together

Put the four definitions side by side and the design reads cleanly:

• The skill.md is the agent-readable document.


• The API it describes is agent-native.


• The self-serve key is how an agent authenticates to it.

The live worked example is ModelVI for Agents. What the public skill.md documents today is the scheduling and posting core: GET /model_list and GET /account_list to see what you can post as, POST /schedule to queue a post, and GET /schedule_result/{id} to check how it went — one API that reaches 14 creator platforms (OnlyFans, Fansly, Fanvue, FanCentro, Maloum, Knky, LoyalFans, MymFans, FetLife, 4Based, BestFans, F2F, FansyMe, and Brezzels) without a separate integration for each. The same partner API also supports more — content import, reposting and story/comment loops, per-org health checks — which is being surfaced as additional agent skills next.

Worth stating plainly, since it's part of an honest definition: an automation layer like this is independent, not affiliated with or endorsed by those platforms, and a well-behaved agent posts only real, manager-approved human captions — never test or bot text.

See one in the wild

The fastest way to make all of this concrete is to look at an actual file. Here's a real skill.md in the wild → Read it top to bottom; it will make every definition above click.

When you want to point your own agent at it, modelvi.com has a "Link my agent" panel that hands you the skill URL and lets you mint a key.