The new Data Agent enhances agent capabilities through an extension system. This topic describes how to install and manage the three extension types (expert suites, skills, and MCP servers) and use them in tasks.
Feature map
Before you begin, we recommend that you understand the functions of and relationships between the three types of objects, as well as the recommended usage order, to configure extensions efficiently.
Object | Role | Relationship with other objects |
Expert suite | A configurable bundle of capabilities | Can contain components such as skills, MCP servers, sub-agents, hooks, context files, and LSP servers |
Skill | An independently manageable capability unit | Built-in skills, independently installed skills, and skills bundled in suites share the same skill list. View, enable, and disable them on the Skills page. |
MCP server | An independently manageable external connection unit that provides tools and data sources | The MCP Servers page manages only independently configured MCP servers. MCP servers built into expert suites are installed with the suite and managed within the suite. They do not appear in this list. |
Recommended usage order: install or enable expert suites as needed first, and then add standalone skills or MCP servers as needed.
Prerequisites
A Data Agent instance is created and started. For more information, see New data agent.
Specific prerequisites for each extension type, such as authentication credentials and naming conventions (Git credentials, OAuth, and scoped package names), are described in the corresponding sections.
Manage expert suites
An expert suite is a configurable bundle of capabilities that can contain components such as skills, MCP servers, sub-agents, hooks, context files. Install suites from different sources to enhance Data Agent.
Install an expert suite
In the left-side navigation pane of Data Agent, click .
On the Expert Suites page, click Install Suite.
In the Install Expert Suite dialog, select an installation method. The following three installation sources are supported:
Remote Repository: Retrieves suite content from a Git repository or a suite marketplace. Use this source for suites hosted in HTTPS Git repositories or published to Claude Marketplace or the Gemini marketplace. Use Ref to pin a branch, tag, or commit.
npm Package: Downloads and installs a versioned suite package from an npm registry. Use this source for suites distributed as versioned npm packages. Append
@versionto the package name to specify a version.Local Path / Upload: Installs from a local directory or an uploaded archive. This source does not depend on a repository or registry, and suits offline environments or unpublished in-house suites.
The main differences between Remote Repository and npm Package are as follows: the former retrieves content from a Git repository or marketplace, pins the version with Ref, and uses the Git credentials in the instance for private repositories; the latter retrieves published package artifacts from an npm registry, pins the version with an
@versionsuffix in the package name, and uses the.npmrc/NPM_TOKENconfiguration of the instance for private registries. Provide the following information for each source:Installation source
Input
Description
Remote Repository
Repository: enter a repository identifier in one of three formats:
owner/repo(for example,myorg/my-suite), an HTTPS Git URL, orowner/repo:plugin-name(to install one plugin from a multi-plugin repository)Ref: optional. Enter a branch, tag, or commit to pin the version. If left empty, the default branch of the repository is used.
Installs from Claude Marketplace, Gemini CLI extensions, or HTTPS Git repositories. You can browse the Gemini and Claude Code marketplaces. If no plugin is specified for a marketplace source, you select one during installation. Private repositories use the Git credentials in the instance.
npm Package
Package name: enter a published scoped package name, optionally with a version, such as
@scope/my-extension@1.2.0Registry: the npm registry to pull from. If left empty, the
.npmrc/NPM_TOKENconfiguration of the instance is used.Installs a scoped package from an npm registry. Only scoped package names (
@scope/name) are supported to avoid confusion with theowner/repoformat.Local Path / Upload
A local directory path or an archive
Installs from a local directory or an uploaded archive
NoteRemote Repository always installs through the daemon. npm Package submits directly when the daemon supports it. Local Path and uploaded archives still run
qwen extensions installin the terminal.Click Confirm Installation to complete the installation. After the installation succeeds, verify on the Expert Suites page that the suite appears in the list. If the installation fails, check the error message on the suite entry.
Manage installed expert suites
After installation, enable or disable a suite from its action menu. Suite enabling and disabling (including bulk operations) take effect within the current instance:
Action | Description |
Enable Suite | Enables the suite within the current instance |
Disable Suite | Disables the suite within the current instance |
Enable All | Enables all suites in the current filtered results in bulk |
Disable All | Disables all suites in the current filtered results in bulk |
Use expert suites in a task
In the Data Agent task input box, enter @ and select Expert Suites from the menu to reference an enabled expert suite. The agent then uses the built-in capabilities of the suite (skills, MCP servers, and more) to complete the task.
The @ menu also references resources such as Files (workspace files), Nodes/Code Files (DataWorks nodes or code files), Tables (data tables), and Data Albums.
Manage skills
A skill is an independently manageable capability unit that falls into two categories: built-in DataWorks skills and independently installed skills. Independently installed skills come from the open skill marketplace, or from uploaded ZIP packages or installation commands. All skills share the same skill list. View, enable and disable them on the Skills page.
Built-in skills
Data Agent ships with built-in DataWorks skills that require no installation. After a built-in skill is enabled, it can be invoked in tasks. On the Skills page, switch to the DataWorks tab to view the built-in skills available in the current instance and enable or disable them as needed. The following table provides examples of frequently used built-in skills. The actual list is subject to the product UI.
Skill name | Description |
DataWorks | The DataWorks assistant supports ad hoc SQL execution, data sampling, and field profiling, and provides complete data service management, covering business processes, APIs, functions, app credentials, authorization, invocation, testing, publishing, and unpublishing. It also provides data security center capabilities, including proactive authorization and approval, data masking, sensitive data identification, and query result control. |
DataWorks data quality inspection scheduling encapsulation | Encapsulates data quality inspection task execution, cross-executor history recovery, event reporting, and inspection report submission. |
DataWorks ChatBI | A DataWorks ChatBI skill for data queries and in-depth analysis. It provides upper-layer agents with support for dataset selection, ChatBI session creation, quick data queries, and in-depth analysis task execution. |
DataWorks Semantic | Pulls, snapshots, and rolls back DataWorks semantic layer YAML artifacts (semantic_model, semantic jobs, and semantic job results), used to synchronize or download semantic layer YAML files or roll back semantic versions. |
DataWorks AI Processing | Generates Python scripts that process videos, images, audio, and documents, and outputs ready-to-run DashScope SDK code. It covers video summarization, frame extraction, analysis, moderation, Q&A, and subtitling; image OCR, description, and object recognition; audio transcription, sentiment analysis, and speaker identification; invoice recognition, PDF parsing, and contract parsing; batch text classification, entity extraction, and embedding; as well as cross-modal pipelines. |
qwen-workflow | Organizes the context of the current task into a task flow YAML file, or matches, modifies, or deletes task flows saved in the current Data Agent workspace and runs them through Qwen Workflow. |
Install a skill
On the Open Skills tab, click Install on a skill card to directly install a marketplace skill. To install a custom skill, use the Install Skill dialog to upload a ZIP package or run an installation command.
In the left-side navigation pane of Data Agent, click .
On the Skills page, click Install Skill.
In the Install Skill dialog, select an installation method and provide the required information:
Upload a ZIP package
Select an installation location:
Installation location
Description
Install to current instance
Writes to the instance locally only and does not upload to the platform. Can update an existing local skill.
Publish to skill platform
Uploads to the cloud and shares within the selected scope, and also installs to the current instance. You must select a scope: Workspace or Tenant.
Click or drag a file into the dashed box to upload a ZIP package. Only
.zipfiles are supported, up to 50 MB per file.ImportantThe ZIP root directory must contain
SKILL.md. The skill name and description are read from theSKILL.mdmetadata, and other files serve as resource attachments referenced by relative path. Limits:SKILL.md≤ 200 KB; the entire ZIP package ≤ 50 MB.The ZIP package supports the following two structures: place
SKILL.mdand resource files directly in the root directory, or place them under a single top-level directory:# Structure 1: root-level structure my-skill.zip ├── SKILL.md # Required. Contains the name, description, and body ├── data/sample.csv # Resource attachment └── images/demo.png # Structure 2: single top-level directory wrapper my-skill.zip └── my-skill/ ├── SKILL.md # Required. Contains the name, description, and body ├── data/sample.csv # Resource attachment └── images/demo.pngClick Install to Instance.
Command install
In the Command Install text box, paste a skill installation command for Qwen Code.
Click Run Install.
Command Install requires public access on the current resource group.
On the Skills page, verify that the target skill appears in the list.
Manage installed skills
On the Skills page, perform the following actions on a target skill:
Action | Description |
Enable Skill | Enables the skill within the current instance |
Disable Skill | Disables the skill within the current instance |
Enable All | Enables all skills in the current filtered results in bulk |
Disable All | Disables all skills in the current filtered results in bulk |
Use skills in a task
In the task input box, enter / to select and invoke a target skill. The agent executes the current task following the predefined flow of the skill.
Manage MCP servers
Model Context Protocol (MCP) servers provide Data Agent with access to external tools and data sources. You can configure multiple MCP servers with different transport protocols. Independently configured MCP servers share the same list as MCP servers built into expert suites. View, enable, and disable them on the MCP Servers page.
Add an MCP server
In the left-side navigation pane of Data Agent, click .
Click Add MCP Server.
In the Add MCP Server dialog, configure the following parameters:
Parameter
Type
Required
Description
Name
Text input
Yes
A unique name for the MCP server
Transport method
Selection
Yes
Streamable HTTP (default), stdio, or SSE. See the following table for the differences.
Input method
Selection
No
Manual Input (default) or JSON Parse. With JSON Parse, you can paste a JSON configuration directly.
Differences among the three transport methods:
Transport method
Connection
Use case
HTTP
Connects to a remote MCP Server over HTTP
MCP Servers deployed on a network or in the cloud
stdio
Starts a local process and communicates through standard input/output
Local command-line MCP Servers (such as npm packages or executables)
SSE
Connects to a remote service over a Server-Sent Events persistent connection
Remote services that still use the SSE protocol
Provide different connection parameters depending on the selected transport method:
HTTP
Parameter
Required
Description
Server endpoint address
Yes
The Streamable HTTP endpoint URL, such as
https://example.com/mcpRequest Headers
No
One KEY=VALUE pair per line, such as
Authorization=Bearer xxxxTimeout
No
Connection timeout in milliseconds. The input box hint is 30000. Enter a value based on your actual network conditions.
stdio
Parameter
Required
Description
Command
Yes
The command-line command to start the service
Arguments
No
Command-line arguments (string array)
Environment Variables
No
Environment variables (key-value pairs)
Timeout
No
Connection timeout in milliseconds
SSE
Parameter
Required
Description
URL
Yes
The SSE endpoint URL
Request Headers
No
One KEY=VALUE pair per line
Timeout
No
Connection timeout in milliseconds
(Optional) Turn on Trust this service. When enabled, tool calls from this server require no confirmation in trusted working directories.
Click OK to create the server. After creation, verify in the MCP Servers list that the status column of the newly added entry shows Connected. If it shows Needs Auth, complete the
OAuthauthentication as instructed. If it shows Error, check the address, transport parameters, and network connectivity, and try again.
To add an MCP server from a JSON configuration, select JSON Parse as the input method and paste the JSON content directly.
Manage existing MCP servers
The following actions are available for configured MCP servers:
Action | Description |
Delete | Deletes the MCP server (confirmation required) |
Enable Service | Enables a disabled service |
Disable Service | Temporarily disables the service within the current instance |
Authenticate | Completes the |
Clear Auth | Clears stored authentication credentials |
Use MCP servers in a task
In the task input box, enter @ and select MCP Resources to reference MCP server resources. The agent invokes the external tools and data sources provided by the MCP server when executing tasks.