Using MCP Resources & Prompts in a Session

Resources via @-mentions

Type @ in the prompt and you'll see resources from connected servers alongside your files. The format is @server:protocol://path:

can you analyze @github:issue://123 and suggest a fix?
compare @postgres:schema://users with @docs:file://database/user-model

Referenced resources are fetched and attached automatically, so Claude starts with the data instead of spending tool calls finding it.

Prompts as slash commands

MCP prompts show up in the / menu as /server:prompt (MCP). You can also type them with the mcp__ form:

/mcp__github__pr_review 456
/mcp__jira__create_issue login-bug high

Arguments are separated by spaces.

Permissions for MCP tools

MCP tools are named mcp__<server>__<tool>, and permission rules use the same names:

{
  "permissions": {
    "allow": ["mcp__sentry__search_issues", "mcp__docs__read_doc"],
    "ask": ["mcp__docs__edit_doc"],
    "deny": ["mcp__github__delete_repository"]
  }
}

mcp__sentry__* covers every tool from one server. Read-only tools are good allow candidates, and anything that writes or deletes should ask or be denied.

MCP + skills: access plus know-how

MCP provides the access. A skill provides the expertise to use it well. For example:

• MCP: a Postgres server with run_query.
• Skill: analytics-queries, which holds your table guide, the metrics definitions finance trusts, and a rule that queries over 30 days must use the daily rollup tables.

Together, Claude can answer "what was churn last quarter?" correctly, not just quickly.

Check your understanding