Does Agent Plugins support mean identical behavior in every AI client?

No. The portable package can carry skills and MCP configuration, but clients differ in supported component types, transports, installation, policy, credentials, extensions, approval UX, and runtime isolation.

Does Data Agent Kit portability grant access to BigQuery or other data services?

No. Packaging and client loading do not grant service authorization. Teams must configure current product prerequisites and least-privilege Google Cloud identities, datasets, projects, regions, network controls, and approvals separately.

AI Agent - Aug 9, 2026

Google Agent Plugins: Agents CLI and Data Agent Kit Guide

Q: What is the difference between Agents CLI and Data Agent Kit?

Agents CLI packages expert skills for agent building and operations. Data Agent Kit packages data-focused skills and MCP servers for Google Data Cloud workflows such as managing assets, running queries, and deploying pipelines.

Quick answer

Google announced Agent Plugins 1.0 support in two product families on August 6, 2026:

Agents CLI packages expert skills for building, evaluating, deploying, observing, and publishing agents.
Data Agent Kit packages data-focused skills and MCP servers for Google Data Cloud workflows, including products such as BigQuery, Spanner, and Cloud SQL.

Google also said it was joining the Agent Plugins core maintainer group. The adoption is about a shared package format. It does not make install commands, permissions, credentials, sandboxing, or behavior identical across AI coding clients.

Why Google adopted the format

The launch post describes a wrapper problem: a portable skill or MCP server still had to be repackaged for clients with different manifests, paths, and configuration shapes. Agent Plugins provides one root plugin.json, fixed skills/ and mcp.json locations, and a namespaced place for client-specific extras.

This reduces package drift without standardizing everything. The portable v1 core remains Skills and MCP servers. Client-specific hooks, commands, agents, policy, and UX remain outside it.

Agents CLI path

Google describes Agents CLI as a way to bring expert agent-building and agent-operations skills into supported AI coding environments. Its skills cover lifecycle work such as building, evaluation, deployment, observability, and publishing.

Use this path when the primary need is agent engineering guidance and workflows. Before installation, inspect the current Agents CLI documentation and package source, then record:

the exact product and package version;
which skills are included and when they trigger;
target client and Agent Skills compatibility;
scripts, external commands, files, network destinations, and data access;
install, update, disable, and removal behavior;
required approvals for deployment or production operations.

A packaged deployment skill is not deployment authority. Keep cloud credentials, project scope, release gates, and rollback owned by the target platform workflow.

Data Agent Kit path

Google describes Data Agent Kit as a collection of plugins that bring Google Data Cloud capabilities to supported coding agents and IDEs. It combines agentic skills with MCP servers for data engineering and development tasks.

Use this path when the agent needs data-platform knowledge plus connected tools. Treat each MCP server as a service boundary. Review:

Google Cloud organization, project, dataset, instance, and Region scope;
user or workload identity and credential source;
read, query, job, mutation, deployment, and administrative permissions separately;
data classification, residency, retention, export, and logging requirements;
query cost, job quotas, timeouts, and cancellation;
confirmation for writes, DDL, pipeline deployment, destructive operations, and cross-project access.

Agent Plugins portability does not grant access to BigQuery, Spanner, Cloud SQL, or any other service. It only standardizes how supported components are packaged and discovered.

Choose by job, not vendor label

Need	Better starting point	Why
Build or evaluate an agent	Agents CLI	Expert skills target agent development and operations
Query or manage Google Data Cloud	Data Agent Kit	Data skills and MCP servers target data workflows
One isolated skill	The skill alone may be enough	A plugin wrapper may not add value
One MCP server in one client	Native MCP configuration may be enough	Avoid packaging overhead when portability is not required
Cross-client bundle of related skills and tools	Agent Plugin package	Common manifest and fixed discovery reduce wrapper drift

Do not assume the two Google products share installation, authentication, runtime, or permissions merely because both use Agent Plugins.

Compatible does not mean identical

The Agent Plugins specification allows a conformant client to support skills, MCP servers, or both. For MCP, clients can differ by stdio and Streamable HTTP support. Namespaced client extensions are intentionally ignored by clients that do not implement them.

Build a tested matrix for each intended host:

package and schema version;
supported component types and MCP transports;
extension namespaces required by the package;
install, enable, update, and removal flow;
credential discovery and secret storage;
tool permission and confirmation model;
process, filesystem, and network isolation;
failure reporting and partial-load behavior.

A vendor launch statement is evidence of adoption, not a permanent compatibility guarantee for every future version or configuration.

What v1 deliberately leaves out

Google’s announcement is explicit that Agent Plugins v1 defines packaging only. It does not define:

an installation mechanism;
a distribution protocol;
a permission model;
sandboxing requirements;
trust or provenance verification;
user experience.

Those controls can differ sensibly between an IDE, CLI, and managed enterprise platform. Teams still need an approved source, integrity checks, least privilege, isolation, audit, lifecycle policy, and rollback.

Adoption workflow

Name the job. Agent engineering, evaluation, data query, asset management, or deployment.
Choose the smallest package. Avoid installing a broad bundle for one narrow task.
Inspect portable and client-specific parts. Review plugin.json, every skill, mcp.json, executables, remote endpoints, and extensions.
Select the target client. Confirm component and transport compatibility from current documentation.
Bind least privilege. Configure product credentials and service authorization outside the package.
Test denied paths. Wrong project, dataset, tenant, user, tool, transport, and destination should fail safely.
Record lifecycle controls. Version pin, updates, disablement, removal, logs, ownership, and rollback.

Read the Agent Plugins 1.0 specification guide to understand the format, compare Agent Plugins, Agent Skills, and MCP before choosing a wrapper, and use the manifest and component authoring guide for your own package.

Frequently asked questions

Which Google products support Agent Plugins 1.0?

Google’s August 6, 2026 announcement identifies Agents CLI and Data Agent Kit and says Google is joining the core maintainer group.

What is the difference between Agents CLI and Data Agent Kit?

Agents CLI packages agent-building and operations skills. Data Agent Kit packages data-focused skills and MCP servers for Google Data Cloud workflows.

Does support mean identical behavior in every AI client?

No. Clients differ in components, transports, policy, credentials, extensions, approval UX, and isolation.

Does Data Agent Kit portability grant data-service access?

No. Teams must separately configure least-privilege identities, resource scope, network controls, and approvals.

Official sources

Source check: August 9, 2026. Verify current product status, installation, supported clients, plugin contents, authentication, Google Cloud permissions, regions, quotas, prices, and data controls before adoption.