Skills
A Skill is MCP Gateway's unit for packaging a task-specific capability instructions for an LLM to follow, bundled with the exact tools and prompts it needs to act on them. You define a Skill once, and MCP Gateway exposes it to any connected agent as a ready-to-use capability, discoverable through the MCP itself.
The instructions follow the community SKILL.md convention that includes a name, a description, and a Markdown body. MCP Gateway extends that convention with a capability group, which is a curated set of tools and prompts drawn from your connected catalog that travels with the instructions as one object. This is the core distinction from a plain SKILL.md file. A Skill in MCP Gateway is not just text for the LLM to read, it is paired with live, callable capabilities from the moment it is discovered.
Why Skills exist
Before Skills, getting an agent ready for a specific job required two separate steps:
- Writing instructions somewhere.
- Configuring which tools it could call through a tool group.
Keeping those two in sync as tasks and tools evolved was manual and error-prone.
Skills replace tool groups. A tool group's curated tool selection is the capability group of a Skill. One object to author and maintain instead of two disconnected ones. MCP Gateway migrates existing tool groups into Skills automatically and irreversibly. No manual re-setup is required.
Anatomy of a Skill
A Skill has the following fields:
| Field | Type | Description |
|---|---|---|
name | string, required | Slug identifier. Lowercase letters, digits, and single hyphens only (^[a-z0-9]+(?:-[a-z0-9]+)*$), maximum 64 characters. Used as the resource identity and, if exposed, the slash-command name. Immutable once the Skill is created. |
description | string, required | Maximum 1024 characters. Shown in the catalog and used by the LLM to decide relevance. |
body | string (Markdown), required | The instruction content itself: what the agent should do and how. |
How a Skill reaches the LLM
Once enabled, MCP Gateway serves a Skill through two independent MCP surfaces at once:
- As a resource:
skill://<id>/SKILL.md. Any MCP client can list and read it. So an agent can discover a relevant Skill on its own during a task, without a person invoking it explicitly. - As a prompt: MCP Gateway registers it as a slash command using the Skill's name (for example,
/support-ticket-triage), for explicit invocation in conversation.
In the same session, the Skill's capability group tools and prompts are live and callable. An agent that pulls in the Skill can immediately call the tools it grants, without any additional setup step.
