Tools & Capabilities

Capability types

AgentBox capabilities fall into several categories:

• Built-in tools: tools the LLM can call while answering, such as Web Search, Image Gen, and SOE.
• Knowledge Base: after association, the system automatically injects retrieval capabilities into the agent.
• Chat features: ASR voice input and TTS voice playback in the chat UI.
• Model capabilities: built-in models and custom models (BYOK).

These capabilities are enabled in different places and have different invocation and billing behavior.

Built-in tools

Web Search

Web Search lets the LLM search for current information when needed. After an admin enables it in the agent's Tools tab, the search tool is injected into that chat request.

Web Search depends on the platform-side search service configuration. Successful invocations are recorded with fixed credits.

Image Gen

Image Gen lets the LLM request an image from a text description. After an admin enables it in the Tools tab, the image generation tool is injected into that chat request.

Image Gen depends on the platform-side image model configuration. Successful invocations are recorded with fixed credits.

SOE

SOE evaluates a user audio file against reference text using Tencent SOE. It returns pronunciation score, accuracy, fluency, completeness, and words that need improvement.

SOE is an LLM tool. It requires the user message to include an audio file key and the model to call the tool at the right moment. Admins can configure the assessment language as Chinese or English.

Knowledge Base

Knowledge Base is not a manually toggled built-in tool. After an admin attaches knowledge bases in the agent's Knowledge Bases tab, AgentBox automatically injects retrieval, document reading, and document listing capabilities into that agent.

The current Knowledge Bases area supports creating knowledge bases, uploading documents, processing documents, and attaching them to agents. Knowledge bases are useful for internal materials, course content, SOPs, FAQs, and similar reference content.

ASR Voice Input

ASR voice input is a chat feature, not an LLM tool. When enabled, users can record audio in chat and have it transcribed to text. Current configuration includes recognition language.

ASR calls use the speech recognition service and are recorded with fixed credits.

TTS Voice Playback

TTS voice playback is a chat feature, not an LLM tool. When enabled, assistant replies can be synthesized as audio. Current configuration includes speed and whether new messages auto-play.

TTS calls use the speech synthesis service and are recorded with fixed credits.

Custom Models (BYOK)

Custom models let admins add their own model provider configuration, including model name, API key, and optional base URL. Agents can then select that custom model.

All workspaces can use custom models. When an agent uses a custom model, model calls do not consume AgentBox credits. If the same agent also uses built-in tools, ASR, or TTS, those platform services are still recorded and charged according to their own usage rules.

Credits lifecycle

Free workspaces receive 100 one-time credits. Essential receives 2,000 credits per month, Standard receives 5,000 credits per month, and Growth receives 20,000 credits per month. Plan credits and extra credits expire at the end of the current billing period and do not roll over.