Agent skills.
Every AI.

An agent skill is a packaged, reusable AI capability — a prompt contract plus an optional backing edge function — that you install into a workspace and call from any AI. Think of it like an npm package, but for AI behaviour rather than JavaScript code. A skill named github-review, for example, bundles the instructions for reviewing a pull request, any tools it needs, and the calling convention your AI follows. Once installed, every agent connected to that workspace can invoke it by name. Skills are the unit of reuse in Aerostack: write the capability once, share it across Claude, ChatGPT, Cursor, and Gemini without rewriting anything per model.

An MCP server wraps an external API — Stripe, GitHub, Notion — and exposes its raw operations as tools. A skill is one level higher: it bundles a specific goal, instructions for how to accomplish it, and optionally a backing edge function into one installable unit. Where an MCP server gives an AI the tools to interact with GitHub, a github-review skill tells the AI how to use those tools to review a pull request well, with consistent behaviour every time. You can compose them: a skill can call tools from MCP servers already installed in your workspace. The practical difference is that MCP servers are API adapters; skills are reusable agent behaviours.

Every AI connected to your Aerostack workspace can call the same installed skills — Claude, Claude Code, ChatGPT, Cursor, and Gemini all work out of the box. This is what cross-model means in practice: you write a skill once and the gateway auto-generates the wire format each ecosystem expects. Claude and Cursor connect over MCP, ChatGPT reads the OpenAI function-calling format from the /openai-tools endpoint, and Gemini reads its function declarations from the /gemini-tools endpoint. You never write an adapter — the same skill definition is exposed in all three formats automatically. If you switch your default model tomorrow, your skills keep working unchanged.

Run aerostack skill install followed by the author and slug — for example, aerostack skill install @johndoe/github-skill — and the skill is immediately live in your workspace. There is nothing to deploy separately. If the skill is backed by an edge function, that function runs on Cloudflare's infrastructure, not yours, so there are no servers to provision or keep warm. You can also install skills directly from the Explore catalog in the dashboard. Private skills require an access token from the publisher; everything else in the public catalog installs with the single command.

Yes. Build a skill with the Aerostack SDK, then publish it as public or private. Public skills appear in the community skills marketplace for anyone to discover and install. Private skills are scoped to your workspace or team and never appear in the public catalog, which makes them suitable for proprietary company capabilities you want to share internally without exposing them to the broader marketplace. You can also gate a public skill with key-required access, so callers need an access token you distribute before they can run it.

A skill can be prompt-only or function-backed — you choose. A prompt-only skill is a curated instruction set: precise, versioned, and consistent across models. A function-backed skill adds a TypeScript edge function that runs on Cloudflare Workers, giving it access to native bindings like D1 databases, R2 storage, KV, and AI APIs without any server setup. The function executes at Cloudflare's edge globally, so the skill runs close to the user with near-zero cold start. This is what separates Aerostack skills from a prompt file you paste into a system prompt: they can actually do work, not just describe how to do it.

The free tier includes 500,000 AI tokens per month across your workspace. Installing community skills costs nothing; usage is metered when the skill executes — tokens consumed by the model interaction and, for function-backed skills, edge function invocations. If you bring your own model API key (OpenAI, Anthropic, Azure, or others), you bypass the platform token pool entirely and pay the provider directly with no markup.

Reach for a skill when you want a reusable AI behaviour that travels with the workspace rather than living in a specific client's system prompt or codebase. If the same capability needs to work across multiple AIs, multiple projects, or multiple teammates, a skill is the right abstraction — install once, available everywhere. If your requirement is truly one-off, a direct system prompt or a quick custom tool is simpler. And if you need to wrap an external API's raw operations without adding any agent-specific instructions, a plain MCP server is the right fit. Skills shine when the value is the opinionated, reusable behaviour layered on top of raw tool access.