Yup. MCP can be configured on a repo level. At task execution time, enabled MCP servers are written as a .mcp.json file into the agent's worktree. Enabled skills are written as .claude/commands/{name}.md files in the worktree, making them available as slash commands to the agent