> ## Documentation Index
> Fetch the complete documentation index at: https://docs.veadk.xyz/llms.txt
> Use this file to discover all available pages before exploring further.

# Built-in tools

`harness.tools` adds built-in tools for one invocation. Its value is a comma-separated string of tool names, such as `"web_search,web_fetch"`. It does not accept tool objects, function arguments, or Python import paths

## Version and scope

This page describes AgentKit CLI **0.54.0** with its default **VeADK 1.0.8** dependency. This version exposes **13 tools** by name, listed below. Every entry is an individual tool, not a toolset containing multiple tools

Request-level `tools` is available on non-shared-OAuth Harness deployments through `POST /harness/invoke` and `POST /run_sse` requests carrying a `harness` configuration. Shared OAuth Harness deployments do not expose general request-level tool, model, or prompt overrides. These fields cannot add temporary tools to those deployments. The deployment's `TOOLS` configuration defines its base tools. If it is unset or empty, none of the tools on this page are loaded by name

Supported names depend on the VeADK version installed on the server. For custom images or manually upgraded dependencies, [check the actual runtime](#check-the-actual-runtime) before using a name

## Supported tools

A registered name does not mean its cloud service, credentials, or sandbox is ready. Configure the environment variables below on the **Harness server**, not inside the HTTP request's `tools` field

| Name | Purpose | Requirements |
| - | - | - |
| `web_search` | Search the web and return summaries of matching pages | Volcengine web search with valid AK/SK or execution-role credentials; see the BytePlus conditions below |
| `parallel_web_search` | Run multiple web searches concurrently and return results by query | Volcengine web search with the same credentials as `web_search`; this version does not switch to the BytePlus search service |
| `web_fetch` | Fetch public pages as Markdown or plain text; also extract PDF text | No additional API key; the target public URL must be reachable, and PDF extraction requires `pypdf`; does not execute JavaScript or access private network addresses |
| `vesearch` | Search the web, social media, and news and return a summary | An enabled search agent, `TOOL_VESEARCH_ENDPOINT`, and valid service credentials; set `TOOL_VESEARCH_API_KEY` explicitly if needed |
| `link_reader` | Read titles and content from web pages, PDFs, or Douyin videos | Volcengine Ark LinkReader access and valid Ark credentials, optionally supplied as `MODEL_AGENT_API_KEY`; at most 3 URLs per call |
| `run_code` | Execute Python or Bash in an AgentKit sandbox and return the result | An available code sandbox, `AGENTKIT_TOOL_ID_SCRIPT`, and sandbox invocation credentials; falls back to `AGENTKIT_TOOL_ID` when the specific ID is unset |
| `coding` | Delegate a coding task to a preconfigured OpenCode sandbox | An AgentKit sandbox configured with OpenCode, `AGENTKIT_TOOL_ID_OPENCODE`, and sandbox invocation credentials; falls back to `AGENTKIT_TOOL_ID` when the specific ID is unset |
| `image_generate` | Generate one or more images from prompts and reference images | Access to an available image-generation model; configure `MODEL_IMAGE_API_KEY`, `MODEL_IMAGE_NAME`, and `MODEL_IMAGE_API_BASE` as needed |
| `image_edit` | Edit source images from prompts, including batch tasks | Access to an available image-editing model; configure `MODEL_EDIT_API_KEY`, `MODEL_EDIT_NAME`, and `MODEL_EDIT_API_BASE` as needed |
| `video_generate` | Generate videos from text, first/last frames, or multimodal references, including batch tasks | Access to an available video-generation model; configure `MODEL_VIDEO_API_KEY`, `MODEL_VIDEO_NAME`, and `MODEL_VIDEO_API_BASE` as needed |
| `text_to_speech` | Synthesize PCM audio from text and return its server-side path | An enabled speech synthesis service, `TOOL_VESPEECH_APP_ID`, and valid speech credentials; set `TOOL_VESPEECH_API_KEY` explicitly if needed; see voice and output settings below |
| `get_city_weather` | Return fixed example weather for predefined cities | No service credentials; use English city names; demonstration data, not live weather |
| `get_location_weather` | Return randomly generated example weather | No service credentials; demonstration data, not live weather |

To read PDFs, include `pypdf` in the Harness deployment dependencies and rebuild the image. Ordinary web-page extraction does not require this PDF parser

### Search and sandbox credentials

Volcengine search prefers `TOOL_WEB_SEARCH_ACCESS_KEY` and `TOOL_WEB_SEARCH_SECRET_KEY`. It can also use `VOLCENGINE_ACCESS_KEY` and `VOLCENGINE_SECRET_KEY`, or the runtime's execution-role credentials

For BytePlus search, set `CLOUD_PROVIDER=byteplus` and `BYTEPLUS_WEB_SEARCH_API_KEY`. In VeADK 1.0.8, `web_search` still requires resolvable AK/SK or execution-role credentials. Set BytePlus AK/SK through `BYTEPLUS_ACCESS_KEY` and `BYTEPLUS_SECRET_KEY`. Do not assume that `parallel_web_search` also supports BytePlus search

`run_code` and `coding` require permission to invoke the target sandbox. Use the runtime's execution role or the appropriate cloud provider's AK/SK. Configure the sandbox region with `AGENTKIT_TOOL_REGION`. With `CLOUD_PROVIDER=byteplus`, the default region is `ap-southeast-1`; the Volcengine default is `cn-beijing`

### Images, video, and speech

When no dedicated API key is configured, image generation, image editing, and video generation reuse `MODEL_AGENT_API_KEY` or VeADK's default model credentials. A valid tool name does not grant access to its generation model. The API URL, model name, and credentials must identify a compatible service. Uploading image results to TOS additionally requires `DATABASE_TOS_BUCKET` and valid TOS access credentials

Set the speech voice with `TOOL_VESPEECH_SPEAKER`, which defaults to `zh_female_vv_uranus_bigtts`. Set the audio output directory with `TOOL_VESPEECH_AUDIO_OUTPUT_PATH`; by default, it uses the server's temporary directory. The returned file path is not a public download URL

<Warning>
  Search, sandbox execution, and media generation can incur cloud resource charges. Configure tool credentials and execution environments for the resources the agent is allowed to access
</Warning>

## Input and merge behavior

| `harness.tools` value | Behavior for this invocation |
| - | - |
| Field omitted | Keep the deployment's existing tools |
| `""`, or only whitespace and commas | Add nothing and keep existing tools |
| `"web_search,web_fetch"` | Add these tools to the existing tools; do not add another tool with the same name |
| `" web_search, web_fetch "` | Trim whitespace around each name before loading |
| Unknown name or incorrect capitalization | Log a warning and skip that name; other loadable tools remain available |
| An array, object, or `null` | Fail request validation because the field requires a string |

Request-level `tools` **adds tools; it does not replace or clear deployment tools**, and it does not persistently change the deployment configuration. A missing Python dependency or credential required during initialization also causes that tool to be skipped with a warning. Tools that load successfully can still fail during execution because of credentials, permissions, or external services

`tools` selects the tools available to the agent. The agent supplies search queries, image tasks, code, and other tool arguments when calling a tool; these are not objects inside `harness.tools`. Making a tool available does not guarantee that the agent calls it on every request

The following request adds two tools that do not require cloud service credentials to one invocation of a non-shared-OAuth Harness. The service still needs a working reasoning-model configuration

```bash lines theme={null}
curl -X POST http://localhost:8000/harness/invoke \
  -H 'Content-Type: application/json' \
  -d '{
    "harness_name": "harness_app",
    "prompt": "Read https://example.com and report the example weather for Beijing",
    "harness": {
      "tools": "web_fetch,get_city_weather"
    },
    "run_agent_request": {
      "user_id": "demo-user",
      "session_id": "tools-demo"
    }
  }'
```

When calling through a cloud gateway, add authentication as required by that gateway. Server-side tool credentials and the HTTP caller's gateway credentials are configured separately

## MCP and skills

| Field | Input | Purpose |
| - | - | - |
| `tools` | Comma-separated string of built-in tool names listed on this page | Load individual tools provided by VeADK |
| `mcp_servers` | List of MCP server configurations | Connect to remote MCP servers and discover and use their tools |
| `skills` | Comma-separated string of skill references | Load skill instructions and resources from a skill-hub path, a skill space, or one skill in a space |

Names of MCP tools, skills, or arbitrary Python functions cannot be used directly in `tools`. Skill references such as `clawhub/owner/skill`, `ss-...`, and `ss-...:s-...` belong in `skills`, not the built-in tool list

The VeADK 1.0.8 name list does not include `video_task_query`, `ppt_generate`, `bash_toolset`, or `run_sandbox_agent`. Availability through another VeADK interface does not make a capability loadable by the same name through this Harness field

## Check the actual runtime

Run the following command **inside the Python environment or container used by Harness** to inspect its installed version and supported names. Running it in a different environment on your computer does not verify the deployed tool list

```bash lines theme={null}
python - <<'PY'
from importlib.metadata import version
from veadk.tools import list_builtin_tools

print("VeADK:", version("veadk-python"))
for name in list_builtin_tools():
    print(name)
PY
```

`list_builtin_tools()` lists names supported by the installed version. It does not validate cloud permissions, credentials, or sandbox availability. If an expected tool is missing, check the version, name, and server-side loading warnings, then verify the tool's runtime requirements