MCP prompts

Skills - reusable, named instructions for models - can be served to AI agents as MCP prompts. An agent lists them with prompts/list, reads one with prompts/get and feeds the instructions to its model, so the same skill that shapes your services' LLM calls also shapes what agents do with your tools.

Assign skills to a gateway

The Skills picker on the wizard's step 01, How do agents connect?, selects which skills an MCP gateway serves - each gateway has its own allow list, and a skill that exists on disk but is not assigned to the gateway is never served through it. In enmasse, the assignment is the skills attribute of an mcp_gateway entry.

A gateway advertises the prompts capability in its initialize response only when it has at least one skill assigned - a gateway with no skills does not advertise it.

Prompt responses

prompts/list answers with the names and descriptions of the assigned skills, paginated with cursors the same way tools/list is - the instructions themselves are not included:

{
    "prompts": [
        {"name": "invoice-analysis", "description": "How to analyze invoices and what to flag"}
    ]
}

prompts/get answers with one skill's full instructions, as the single message of the prompt:

{
    "description": "How to analyze invoices and what to flag",
    "messages": [
        {"role": "user", "content": {"type": "text", "text": "..the skill's instructions.."}}
    ]
}

Editing a skill takes effect on the next prompts/get - skill files are read live, never cached.

A name outside the gateway's allow list, or one whose file is gone, is refused with an invalid-params error saying Prompt not found - whether the skill exists elsewhere is never revealed.

With the audit log on, prompts/list and prompts/get are recorded as mcp-prompts-list and mcp-prompts-get events, with the prompt's name in the event's tool column.

See also

Learn more