Agent Registry

Google Cloud Agent Registry Skill

Help users interact with the gcloud alpha agent-registry API by translating natural language into the correct gcloud command, showing it for approval, then executing it.

Workflow

1. Resolve session context — at the start of each session, silently run:

   gcloud config get-value project 2>/dev/null
   gcloud config get-value compute/region 2>/dev/null
   

   Store the results as SESSION_PROJECT and SESSION_LOCATION. Use these as defaults.

2. Parse the user's request to identify the resource and action.

3. Construct the gcloud command using session defaults.

4. Approval:

   • For list and describe commands: Skip explicit approval and execute immediately.
   • For create, delete, or update commands: Show the command and ask for approval: "Ready to run this command? (yes/no)"

5. Execute and display the output.

Auth & Setup

# Check auth
gcloud auth list

# Login
gcloud auth login

# Set project
gcloud config set project PROJECT_ID

# (Optional) Set API override if needed
gcloud config set api_endpoint_overrides/agentregistry https://agentregistry.googleapis.com/

IAM Permissions

Resource Types & Registration

The primary command for registering resources is gcloud alpha agent-registry services create.

1. MCP Servers

Used to register Model Context Protocol servers.

# Register an MCP Server (example: GitHub)
# Ask user to paste contents of mcp-spec.json for --mcp-server-spec-content
gcloud alpha agent-registry services create github \
  --location=us-central1 \
  --display-name="GitHub MCP Server" \
  --description="Connects to GitHub" \
  --mcp-server-spec-type=tool-spec \
  --mcp-server-spec-content='PASTE_MCP_SPEC_JSON_HERE' \
  --interfaces='[{"protocolBinding": "jsonrpc", "url": "https://api.github.com/mcp"}]'

# List MCP Servers
gcloud alpha agent-registry mcp-servers list --location=us-central1

# Filter MCP Servers by Runtime
gcloud alpha agent-registry mcp-servers list \
  --location=us-central1 \
  --filter="attributes.\"agentregistry.googleapis.com/system/RuntimeReference\".uri:reasoningEngine"

# List Global MCP Servers
gcloud alpha agent-registry mcp-servers list --location=global

2. Agents

Used to register AI agents (e.g., A2A, Salesforce).

# Register an Agent (example: Salesforce)
gcloud alpha agent-registry services create salesforce \
  --location=us-central1 \
  --display-name="Salesforce Agent" \
  --description="Salesforce Einstein Agent" \
  --agent-spec-type=no-spec \
  --interfaces='[{"protocolBinding": "http-json", "url": "https://api.salesforce.com/agent/v1"}]'

# Register an A2A Agent (Special Case)
# Ask user to paste contents of agent_card.json for --agent-spec-content
gcloud alpha agent-registry services create testa2a \
  --location=us-central1 \
  --display-name="Test A2A Agent" \
  --description="Sample A2A Agent" \
  --agent-spec-type=a2a-agent-card \
  --agent-spec-content='PASTE_AGENT_CARD_JSON_HERE'

# List Agents
gcloud alpha agent-registry agents list --location=us-central1

# List Global Agents
gcloud alpha agent-registry agents list --location=global

3. Endpoints

Used to register service endpoints (e.g., Vertex AI models).

# Register an Endpoint (example: Gemini Models)
gcloud alpha agent-registry services create gemini-models \
  --location=us-central1 \
  --display-name="Vertex AI Model Garden" \
  --description="List of all models in Vertex AI Model Garden" \
  --endpoint-spec-type=no-spec \
  --interfaces='[{"protocolBinding": "jsonrpc", "url": "https://us-central1-aiplatform.googleapis.com/v1beta1/publishers/*/models"}]'

# List Endpoints
gcloud alpha agent-registry endpoints list --location=us-central1

# Update Endpoint Display Name
gcloud alpha agent-registry services update gemini-models \
  --display-name="Model Garden on Vertex AI" \
  --location=us-central1

4. GKE Based Agents

Used to configure Kubernetes deployments to be registered as an Agent or MCP Server. You can add the required apphub.cloud.google.com/functional-type annotation to YAML files using the provided python script.

# Ask the user if they want to process the current directory or a specific directory/file
# Ask the user if the functional type is an AGENT or MCP_SERVER

# Run the python script to annotate the YAML files
./scripts/annotate_gke.py /path/to/folder_or_file.yaml --type AGENT

5. Bindings

Bindings connect a source agent to a target resource.

• Create a binding: gcloud alpha agent-registry bindings create [BINDING_ID] --agent=[AGENT_ID] --target=[TARGET_RESOURCE_ID] --location=[LOCATION]
• Create a binding with delegated permissions: gcloud alpha agent-registry bindings create [BINDING_ID] --agent=[AGENT_ID] --target=[TARGET_RESOURCE_ID] --auth-provider=[AUTH_PROVIDER_ID] --location=[LOCATION]
• List bindings: gcloud alpha agent-registry bindings list --location=[LOCATION]
• View binding details: gcloud alpha agent-registry bindings describe [BINDING_ID] --location=[LOCATION]
• Update a binding: gcloud alpha agent-registry bindings update [BINDING_ID] --target=[NEW_TARGET_ID] --location=[LOCATION]
• Delete a binding: gcloud alpha agent-registry bindings delete [BINDING_ID] --location=[LOCATION]

Agent Dashboard

The Agent Dashboard provides a consolidated view of all agents in the current project, searching across both global and the regional location (default: us-central1).

To generate the dashboard, run:

./scripts/agent-dashboard.sh

The output will be a Markdown table containing the following fields:

• Name: The ID of the agent.
• Display Name: The human-readable name.
• Location: The region where the agent is registered.
• Runtime: The reference to the agent's runtime.

MCP Server Dashboard

The MCP Server Dashboard provides a consolidated view of all MCP servers in the current project, searching across both global and the regional location (default: us-central1).

To generate the dashboard, run:

./scripts/mcp-dashboard.sh

The output will be a Markdown table containing the following fields:

• Name: The ID of the MCP server.
• Display Name: The human-readable name.
• Location: The region where it is registered.
• Tools: List of tools provided by the server.
• Runtime: The reference to the runtime.

Detailed Command Reference

All commands support --location (required) and --project (optional).

Service Creation Flags

Natural Language → Command Examples

Advanced Filtering

To filter resources based on nested attributes with special characters (like dots or slashes), use double quotes around the key segments in the --filter flag.

[!WARNING] The double-quote escaping shown below (\") works in bash/zsh. Windows CMD or PowerShell users may need different escaping (e.g., " or ``` " ``) for nested attribute keys.

Filtering by agentId and mcpServerId: The agentId and mcpServerId fields uniquely identify an agent or MCP server and follow the URN model. For agents, agentId always begins with: urn:agent:projects-{project-number}:projects:{project-number}:locations:{location}:{other-segments} For MCP servers, mcpServerId always begins with: urn:mcp:projects-{project-number}:projects:{project-number}:locations:{location}:{other-segments} The {other-segments} can vary based on the platform, for example:

• Reasoning Engine: aiplatform:reasoningEngines:{reasoning-engine-id}
• GKE: container:clusters:{cluster-name}:k8s:namespaces:{namespace}:apps:deployments:{deployment-id} You can filter agents and MCP servers by substring matching these fields. For example, --filter="agentId:reason" finds Reasoning Engine agents, and --filter="mcpServerId:run" finds Cloud Run MCP servers.

Mapping Tips:

• Map "runtime" to attributes."agentregistry.googleapis.com/system/RuntimeReference".uri.
• Map "identity" to attributes."agentregistry.googleapis.com/system/RuntimeIdentity".principal.
• Map "tool name" or "tool" to tools.name for MCP Server list commands.
• Synonyms: "agent engine", "reasoning engine", and "vertex ai" all refer to the runtime value reasoningEngine.
• Context Filtering: If the user asks about "agents", use the agents resource group (e.g., gcloud alpha agent-registry agents list), not mcp-servers or endpoints.

# Example: Show all agents where the runtime is reasoningEngine
gcloud alpha agent-registry agents list \
  --location=us-central1 \
  --filter="attributes.\"agentregistry.googleapis.com/system/RuntimeReference\".uri:reasoningEngine"

# Example: Show agents where identity contains a specific service account ID
gcloud alpha agent-registry agents list \
  --location=us-central1 \
  --filter="attributes.\"agentregistry.googleapis.com/system/RuntimeIdentity\".principal:service-432423"

# Example: Search for all reasoning engine agents using the agentId
gcloud alpha agent-registry agents list \
  --location=us-central1 \
  --filter="agentId:reason"

# Example: Search for cloud run MCP servers using the mcpServerId
gcloud alpha agent-registry mcp-servers list \
  --location=us-central1 \
  --filter="mcpServerId:run"

Searching Agents

When searching for agents, always default to using the mcp_agentregistry_search_agents MCP tool if the agentregistry MCP server is available. Use the gcloud alpha agent-registry agents search command as a fallback.

MCP Tool Usage (mcp_agentregistry_search_agents)

• parent: projects/{project}/locations/{location}
• searchString: Follows the same syntax as the gcloud command (e.g., skills.name:model, displayName:Assessor*).

gcloud Fallback Usage

You can search for agents using the gcloud alpha agent-registry agents search command. This command supports a --search-string flag with specific match types:

• Exact match (=): Matches the entire value exactly. Example: --search-string="agentId=\"urn:agent:projects-123:projects:123:locations:us-central1:agentregistry:services:my-agent\""
• Token match (:): Matches individual words exactly. Example: --search-string="agentId:\"urn:agent:projects-123\""
• Prefix match (*): Matches values that start with the given prefix. Example: --search-string="agentId:\"urn:agent:projects-123*\""

Searchable Fields & Examples

You can search across fields such as agentId, displayName, skills.name, skills.id, and skills.description. Note that values containing colons (like URNs) must be escaped.

# Search for agents by agentId (urn) with prefix match
gcloud alpha agent-registry agents search --location=us-east4 \
  --search-string="agentId:\"urn:agent:projects-1064111708665*\""

# Search for agents containing the display name 'Assessor'
gcloud alpha agent-registry agents search --location=us-east4 \
  --search-string="displayName:Assessor*"

# Search for agents by skill name
gcloud alpha agent-registry agents search --location=us-east4 \
  --search-string="skills.name:model"

# Search for agents by skill id
gcloud alpha agent-registry agents search --location=us-east4 \
  --search-string="skills.id:AssessorAgent"

# Search for agents by skill description
gcloud alpha agent-registry agents search --location=us-east4 \
  --search-string="skills.description:severity*"

Searching MCP Servers

When searching for MCP servers, always default to using the mcp_agentregistry_search_mcp_servers MCP tool if the agentregistry MCP server is available. Use the gcloud alpha agent-registry mcp-servers search command as a fallback.

MCP Tool Usage (mcp_agentregistry_search_mcp_servers)

• parent: projects/{project}/locations/{location}
• searchString: Follows the same syntax as the gcloud command (e.g., displayName:GitHub*, mcpServerId:run).

gcloud Fallback Usage

You can search for MCP servers using the gcloud alpha agent-registry mcp-servers search command. This command supports a --search-string flag with the same match types as agents (exact =, token :, prefix *).

Searchable Fields & Examples

You can search across fields such as mcpServerId, name, and displayName. Note that values containing colons (like URNs) must be escaped.

# Search for MCP servers by mcpServerId with prefix match
gcloud alpha agent-registry mcp-servers search --location=us-east4 \
  --search-string="mcpServerId:\"urn:mcp:projects-1064111708665*\""

# Search for MCP servers containing the display name 'GitHub'
gcloud alpha agent-registry mcp-servers search --location=us-east4 \
  --search-string="displayName:GitHub*"

Python ADK Integration

The Google Agent Development Kit (ADK) allows seamless integration with the Agent Registry.

1. Requirements & Setup

• ADK Version: 1.26.0 or higher (minimum).
• Installation:

  # Using pip
  pip install --upgrade google-adk
  
  # Using uv
  uv add google-adk
  

2. Import & Usage

Add the following import to your Python code:

from google.adk.integrations.agent_registry import AgentRegistry

3. Invoking an MCP Server from Registry

Use this snippet to retrieve and use an MCP toolset:

import os
from google.adk.integrations.agent_registry import AgentRegistry

# Initialize registry client
registry = AgentRegistry(project_id=SESSION_PROJECT, location=SESSION_LOCATION)

# Retrieve MCP Toolset using the full resource name
# Example resource name: projects/PRJ/locations/LOC/mcpServers/SERVER_NAME
mcp_toolset = registry.get_mcp_toolset(
    f"projects/{SESSION_PROJECT}/locations/{SESSION_LOCATION}/mcpServers/{MCP_SERVER_NAME}"
)

4. Integrating a Remote A2A Agent

Use this snippet to use a registry agent as a sub-agent:

from google.adk import Agent, Gemini, types
from google.adk.integrations.agent_registry import AgentRegistry

# Initialize registry client
registry = AgentRegistry(project_id=SESSION_PROJECT, location=SESSION_LOCATION)

# Retrieve Remote A2A Agent
remote_agent = registry.get_remote_a2a_agent(
    f"projects/{SESSION_PROJECT}/locations/{SESSION_LOCATION}/agents/{AGENT_NAME}"
)

# Define a new Agent with the remote agent as a sub-agent
help_agent = Agent(
    name="help_agent",
    description="Helpful AI Assistant that uses a remote agent.",
    model=Gemini(model="gemini-2.5-flash"),
    sub_agents=[remote_agent]
)

Interactive Prompts

Only ask if still missing after checking session context:

• location: "Which region? (e.g. us-central1)" — only if compute/region was not set
• project: "Which project?" — only if project was not set in gcloud config
• A2A Agent Card: For A2A agents, explicitly ask: "Please paste the contents of your agent_card.json file." and use it for --agent-spec-content.
• MCP Server Spec: For MCP servers, explicitly ask: "Please paste the contents of your MCP server spec JSON file." and use it for --mcp-server-spec-content.
• GKE Deployments: Ask for the target path (current folder, specific folder, or file) and the functional type (AGENT or MCP_SERVER).

Only ask for what's strictly needed — don't overwhelm the user.

Error Handling

If a command fails:

1. Check if gcloud alpha component is installed.
   • Required (minimum): Google Cloud SDK 565.0.0 or higher
   • Required (minimum): alpha component 2026.04.10 or higher
2. Verify the --location (some resources may be in global or specific regions).
3. Ensure JSON payloads for --interfaces or specs are correctly quoted for the shell.
4. Check project permissions for agentregistry.googleapis.com.

Bug Reporting

If you encounter an unexpected problem, bug, or a failure that you cannot resolve:

1. Ask the user if they would like to create a GitHub issue for this bug.
2. If the user agrees, generate a descriptive title and body for the issue based on the error context.
3. Show the user the proposed issue content and the command to create it.
4. Ask for final approval before running the command.
5. Once approved, use the gh CLI to create the issue in the repository. For example:

   gh issue create --repo agentskills/agent-registry-skill --title "Title of the bug" --body "Description of the bug, including error messages and steps to reproduce."
   

ADK Reference

If the user asks about ADK, read references/adk-docs.md for instructions on which URL to fetch.