Tools

VeADK Java builds on the ADK Java tool system and adds Volcengine-oriented built-in tools and toolsets. You can attach Java functions, web search, web fetch, MCP Router, Skills, Skills Sandbox, Code Sandbox, and other capabilities to the same Agent, then let the model decide which tool to call at runtime.

Entry Point

Tools are usually passed through Agent.builder().tools(...). You can mix BaseTool, Toolsets, and ADK Java FunctionTool instances.

import com.volcengine.veadk.Agent;
import com.volcengine.veadk.Runner;
import com.volcengine.veadk.tools.builtin.mcp.McpRouterToolset;
import com.volcengine.veadk.tools.builtin.webfetch.WebFetchTool;
import com.volcengine.veadk.tools.builtin.websearch.ParallelWebSearchTool;

Agent agent = Agent.builder()
        .name("research_assistant")
        .modelName("doubao-seed-2-1-pro-260628")
        .instruction("Use tools to collect fresh context before answering, and cite the source.")
        .tools(
                new WebFetchTool(),
                new ParallelWebSearchTool(),
                McpRouterToolset.create())
        .build();

try {
    String answer = new Runner(agent).run("Research VeADK Java tool capabilities.");
    System.out.println(answer);
} finally {
    agent.close().blockingAwait();
}

When an Agent uses MCP or remote Toolsets, call agent.close().blockingAwait() after the task so the underlying sessions and connections can be released.

Built-In Tools

Scenario	Java entry point	Tool name	Typical use
Web Search	`new WebSearchTool()`	`web_search`	Run one real-time web search query
Parallel Web Search	`new ParallelWebSearchTool()`	`parallel_web_search`	Run multiple search queries in parallel
Web Fetch	`new WebFetchTool()`	`web_fetch`	Read public web pages or PDF content
MCP Router	`McpRouterToolset.create(...)`	Returned by remote MCP	Reuse centrally managed platform tools
Skills	`new SkillToolset(...)`	Returned by the Skill definition	Expose local or remote Skills as callable tools
Skills Sandbox	`new ExecuteSkillsTool()`	`execute_skills`	Delegate a full task to a remote Skills Sandbox
Async Skills Sandbox	`new InvokeSkillTool()`, `new PollSkillTool()`	`invoke_skill`, `poll_skill`	Create a remote task and poll for the result
Code Sandbox	`new RunCodeTool()`	`run_code`	Execute code in AgentKit Code Sandbox
Code Sandbox Toolset	`CodeSandboxToolset.create(...)`	Returned by Sandbox MCP	Connect to an existing Code Sandbox MCP service
Sandbox session client	`new SandboxClient()`	Not a model tool	Create sessions, upload files, run code, and read outputs in application code
TTS	`new TextToSpeechTool()`	`text_to_speech`	Convert text to speech
PowerPoint	`new PptGenerateTool()`	`ppt_generate`	Generate presentation files
Lark	`LarkToolset.create()`	Returned by Lark MCP	Connect to Lark MCP tools
Mobile Use	`new MobileUseTool(systemPrompt)`	`mobile_use`	Operate a cloud phone environment
Vanna + ClickHouse	`new VannaToolset(...)`	`run_sql`, `summarize_data`, `visualize_data`	Query, summarize, and visualize structured data

Web Search And Web Fetch

Use WebSearchTool for one query and ParallelWebSearchTool when a question should be split into multiple concurrent queries. Use WebFetchTool when the user gives an explicit URL and the Agent should read the page or PDF body.

import com.volcengine.veadk.Agent;
import com.volcengine.veadk.tools.builtin.webfetch.WebFetchTool;
import com.volcengine.veadk.tools.builtin.websearch.ParallelWebSearchTool;
import com.volcengine.veadk.tools.builtin.websearch.WebSearchTool;

Agent agent = Agent.builder()
        .name("web_researcher")
        .modelName("doubao-seed-2-1-pro-260628")
        .instruction("""
                Use web_search or parallel_web_search for fresh information.
                When the user provides a link, call web_fetch first and answer from the content.
                """)
        .tools(
                new WebSearchTool(),
                new ParallelWebSearchTool(),
                new WebFetchTool())
        .build();

web_fetch accepts url, extract_mode, and max_chars. extract_mode defaults to markdown and also supports text. The tool blocks private, loopback, link-local, multicast, and reserved addresses, limits redirects and response size, and is intended for public resources rather than authenticated back-office pages.

MCP Router

If tools are centrally registered in MCP Router, expose the Router as a Toolset. The default constructor reads environment variables, and the explicit constructor lets you pass the URL, API key, and allowed tool names.

import com.volcengine.veadk.Agent;
import com.volcengine.veadk.tools.builtin.mcp.McpRouterToolset;
import java.util.List;

Agent agent = Agent.builder()
        .name("mcp_agent")
        .modelName("doubao-seed-2-1-pro-260628")
        .instruction("Use authorized MCP Router tools for external system requests.")
        .tools(McpRouterToolset.create(
                System.getenv("TOOL_MCP_ROUTER_URL"),
                System.getenv("TOOL_MCP_ROUTER_API_KEY"),
                List.of("weather.query", "calendar.search")))
        .build();

Without arguments, McpRouterToolset.create() reads TOOL_MCP_ROUTER_URL and optional TOOL_MCP_ROUTER_API_KEY. When an API key is configured, the client sends it as Bearer authentication.

Skills

There are two common ways to use Skills. The first is to expose a Skill Source as ordinary tools, so the model can call skills just like any other tool.

import com.google.adk.tools.skills.SkillToolset;
import com.volcengine.veadk.Agent;
import com.volcengine.veadk.skills.VeSkillSource;

String skillSourceId = System.getenv("SKILL_SOURCE_ID");

Agent agent = Agent.builder()
        .name("skill_agent")
        .modelName("doubao-seed-2-1-pro-260628")
        .instruction("Select the right skill for each user request.")
        .tools(new SkillToolset(new VeSkillSource(skillSourceId)))
        .build();

The second is to execute remote skill workflows through Skills Sandbox. For synchronous workflows, use ExecuteSkillsTool; it sends workflow_prompt to the remote skill space and waits for the result.

import com.volcengine.veadk.Agent;
import com.volcengine.veadk.tools.builtin.sandbox.ExecuteSkillsTool;

Agent agent = Agent.builder()
        .name("remote_skill_agent")
        .modelName("doubao-seed-2-1-pro-260628")
        .instruction("When skills are needed, call execute_skills with the full user request.")
        .skills(System.getenv("SKILL_SPACE_ID"))
        .skillsMode("skills_sandbox")
        .tools(new ExecuteSkillsTool())
        .build();

For asynchronous workflows, combine InvokeSkillTool and PollSkillTool. The model calls invoke_skill to create the remote task, then calls poll_skill with the returned task_id until it receives a final result.

import com.volcengine.veadk.Agent;
import com.volcengine.veadk.tools.builtin.sandbox.InvokeSkillTool;
import com.volcengine.veadk.tools.builtin.sandbox.PollSkillTool;

Agent agent = Agent.builder()
        .name("async_skill_agent")
        .modelName("doubao-seed-2-1-pro-260628")
        .instruction("For remote tasks, call invoke_skill first, then poll_skill until done.")
        .tools(new InvokeSkillTool(), new PollSkillTool())
        .build();

Code Sandbox

RunCodeTool lets the model execute code in AgentKit Code Sandbox. It is useful for data processing, temporary code checks, and artifact generation. Tool arguments include code, language, and optional timeout.

import com.volcengine.veadk.Agent;
import com.volcengine.veadk.tools.builtin.sandbox.RunCodeTool;

Agent agent = Agent.builder()
        .name("code_agent")
        .modelName("doubao-seed-2-1-pro-260628")
        .instruction("Use run_code for calculations, file parsing, and small generated outputs.")
        .tools(new RunCodeTool())
        .build();

If you already have a Code Sandbox MCP service, connect it through CodeSandboxToolset.

import com.volcengine.veadk.Agent;
import com.volcengine.veadk.tools.builtin.sandbox.CodeSandboxToolset;

Agent agent = Agent.builder()
        .name("sandbox_mcp_agent")
        .modelName("doubao-seed-2-1-pro-260628")
        .tools(CodeSandboxToolset.create(System.getenv("TOOL_CODE_SANDBOX_URL")))
        .build();

Application code can manage remote sandbox sessions directly with SandboxClient. Paths are resolved under remote /workspace, and the client rejects traversal and unsafe absolute paths.

import com.volcengine.veadk.tools.builtin.sandbox.SandboxClient;

try (SandboxClient client = new SandboxClient()) {
    var session = client.createSession("order-analysis", 1800);
    session.writeFile("input/orders.csv", csvBytes);
    String result = session.runCode(
            "import pandas as pd; print(pd.read_csv('/workspace/input/orders.csv').sum())",
            "python3",
            60);
    byte[] report = session.readFile("output/report.json");
    session.close();
}

Environment Variables

Capability	Environment variables
Volcengine authentication	`VOLCENGINE_ACCESS_KEY`, `VOLCENGINE_SECRET_KEY`, optional `VOLCENGINE_SESSION_TOKEN`
Web Search	`WebSearchTool` uses `VOLCENGINE_ACCESS_KEY` and `VOLCENGINE_SECRET_KEY`. `ParallelWebSearchTool` can use `TOOL_WEB_SEARCH_ACCESS_KEY` / `TOOL_WEB_SEARCH_SECRET_KEY`, falling back to Volcengine AK/SK.
MCP Router	`TOOL_MCP_ROUTER_URL`, optional `TOOL_MCP_ROUTER_API_KEY`
Skills	`SKILL_SOURCE_ID` or `SKILL_SPACE_ID`, plus skill-space policy configuration
AgentKit Code Sandbox	`AGENTKIT_TOOL_ID`, AgentKit host/region, and Volcengine AK/SK
Code Sandbox MCP	`TOOL_CODE_SANDBOX_URL`
TTS	`TOOL_VESPEECH_APP_ID`, `TOOL_VESPEECH_API_KEY`, optional `TOOL_VESPEECH_SPEAKER`
Lark	`TOOL_LARK_ENDPOINT`, `TOOL_LARK_API_KEY`, and `TOOL_LARK_TOKEN`
Mobile Use	`TOOL_MOBILE_USE_TOOL_ID` and Volcengine AK/SK

Automatic Tools

When an Agent is configured with a knowledge base or long-term memory, VeADK Java automatically adds retrieval tools such as loadKnowledgebase and loadMemory. When local Skills are configured and no explicit SkillToolset is provided, the framework also prepares the corresponding Skill tools. For skills_sandbox mode, pass ExecuteSkillsTool explicitly.

Security Notes

Tools let the model access external systems, execute code, and trigger business workflows, so configure credentials and allowed tool names with least privilege. For tools that can write data, approve requests, make payments, delete resources, publish content, or send notifications, add business-level confirmation, auditing, and idempotency protection. Code Sandbox and Skills Sandbox provide execution isolation, but you should still limit input files, timeouts, network access, and the secret scope available to the execution environment.

Overview

Get Started

Core Components

Observability

Deploy

Entry Point

Built-In Tools

Web Search And Web Fetch

MCP Router

Skills

Code Sandbox

Environment Variables

Automatic Tools

Security Notes

​Entry Point

​Built-In Tools

​Web Search And Web Fetch

​MCP Router

​Skills

​Code Sandbox

​Environment Variables

​Automatic Tools

​Security Notes

Entry Point

Built-In Tools

Web Search And Web Fetch

MCP Router

Skills

Code Sandbox

Environment Variables

Automatic Tools

Security Notes