What is an agent skill?
An agent skill is a declarative instruction package, typically a markdown file, that teaches an AI coding agent how to install, authenticate, and use a specific tool. Skills provide knowledge: when to call a tool, what commands exist, and how to structure output. They are distinct from the tools themselves, which provide execution. An agent reads a skill once and applies it across all subsequent tasks that require that capability, without needing the knowledge re-explained each session.
| Factor | Agent skill | MCP server | Direct API integration |
|---|---|---|---|
| What it provides | Knowledge (instructions) | Tool access (execution) | Data (responses) |
| Setup by agent | Automatic on install | Manual configuration | Manual wrapper code |
| Updates | Skill file update | Server redeployment | Code change |
| Context cost | Loaded on demand | Schema loaded upfront | None |
| Best for | Teaching agents to use CLIs and tools | Structured tool calls from the agent | Programmatic pipelines |
Skills are most useful when you want an agent to use a CLI tool autonomously: instead of configuring a tool wrapper manually for every agent harness, you install a skill once and the agent learns the tool's interface, authentication flow, and output conventions on its own. This is particularly valuable for web scraping workflows where the agent needs to decide when to scrape, search, or crawl based on the task rather than executing a hardcoded sequence of calls. For a deeper look at how agent skills, MCP servers, and CLI tools compare on token cost, reliability, and use case fit, see the MCP vs CLI guide for AI agents.
Firecrawl's agent skill installs with a single command and is compatible with Claude Code, OpenAI Codex, Gemini CLI, and other coding agent harnesses. Once installed, the agent knows when and how to use Firecrawl's scrape, search, crawl, and browser commands without further configuration.