ADK Web Documentation

ADK Web is a built-in developer UI integrated with the Google Agent Development Kit (ADK) to facilitate agent development and debugging. Key features visible in the UI include:

• Events: Monitoring agent events.
• Tracing: Debugging execution flows.
• Artifacts: Inspecting generated outputs or files.
• Evaluations: Running and viewing agent performance evaluations.
• Agent Builder & Assistant: Tools for constructing and interacting with agents.

Before running ADK Web, ensure you have the following installed on your system:

• npm
• NodeJs
• Angular CLI
• google-adk (Python)
• google-adk (Java)

You must also clone the adk-web repository to your local machine.

To run the ADK Web developer UI, follow these steps from the root of your local adk-web folder:

1. Install dependencies: Run npm install (use sudo if required by your environment).
2. Start the Web UI: Run the serve script, specifying the backend URL using the --backend flag.
3. Start the API Server: In a separate terminal, run the adk api_server command. You must configure --allow_origins to match the Web UI's address (typically http://localhost:4200) and set the --host to 0.0.0.0 to allow connections.

Once both are running, access the UI at http://localhost:4200.

# 1. Install dependencies
sudo npm install

# 2. Run adk web
npm run serve --backend=http://localhost:8000

# 3. Run adk api server (in a separate terminal)
adk api_server --allow_origins=http://localhost:4200 --host=0.0.0.0

To synthesize simulated user turns using audio, provide a UserSimulatorConfig of type llm_audio. This configuration allows you to specify an audio_model (such as a Gemini TTS model or cloud_tts) and detailed speech_config for voice and language selection.

Default Audio Model: gemini-3.1-flash-tts-preview

Available Prebuilt Voices:

• Kore
• Puck
• Charon
• Aoede
• Fenrir

import { UserSimulatorConfig } from './path/to/Eval';

const config: UserSimulatorConfig = {
  type: 'llm_audio',
  audio_model: 'gemini-3.1-flash-tts-preview',
  audio_model_configuration: {
    response_modalities: ['AUDIO'],
    speech_config: {
      voice_config: {
        prebuilt_voice_config: {
          voice_name: 'Kore'
        }
      },
      language_code: 'en-US'
    }
  }
};

When triggering an evaluation run, you can configure the run mode and the specific metrics to be used.

Run Modes

• standard: The default non-live run.
• live: Uses bidirectional streaming inference.

Metrics Configuration

Evaluations use a set of metrics, each with a configurable threshold. Common metrics include:

• tool_trajectory_avg_score (Default threshold: 1.0)
• response_match_score (Default threshold: 0.7)

Live Audio Simulation

If live mode is enabled and the input modality is set to audio, you can provide a UserSimulatorConfig. This configures an LLM-based audio user simulator (using the llm_audio type) with specific voice and language settings.

The adk api_server command is used to launch the backend required by the ADK Web UI. It accepts the following flags:

• --allow_origins: Specifies the allowed origins for CORS (e.g., http://localhost:4200).
• --host: Specifies the host address to bind to (e.g., 0.0.0.0).

If you encounter adk command not found, ensure google-adk is installed or that your Python virtual environment is activated.

adk api_server --allow_origins=http://localhost:4200 --host=0.0.0.0

The following interfaces define the structure of evaluation data used within the ADK Web UI:

• EvalCase: A single unit of evaluation containing a sessionInput, an optional conversation (array of Invocation), and events.
• EvalSet: A collection of EvalCase objects grouped under an evalSetId.
• Invocation: A single turn in a conversation, containing userContent, an optional finalResponse, intermediateData, and a creationTimestamp.
• EvaluationResult: The output of an evaluation run, linking a setId and evalId to a finalEvalStatus and specific evalMetricResults.

The following types define the operational modes for an evaluation:

• EvalRunMode: 'standard' | 'live'
• LiveInputModality: 'audio' | 'text'

The EvalStatus enum represents the verdict returned by the backend for an evaluation run. Use these values to interpret finalEvalStatus in an EvaluationResult.

// EvalStatus values:
// PASSED = 1
// FAILED = 2
// NOT_EVALUATED = 3

The RunEvalDialogResult interface represents the configuration object emitted when an evaluation configuration dialog is confirmed. This object is consumed by the runEval function to execute the evaluation.

Fields:

• metrics: An array of EvalMetric objects selected by the user, including their respective threshold values.
• useLive: A boolean indicating if the run should use live (bidirectional streaming) mode.
• userSimulatorConfig (optional): Present only when the audio user simulator is enabled for a live run. It contains the UserSimulatorConfig for the audio model, voice, and language.

An EvalMetric defines a specific metric used to score an agent's performance, paired with a passing threshold. You can use the DEFAULT_EVAL_METRICS as a starting point for common evaluations.

import { EvalMetric } from './path/to/Eval';

const myMetric: EvalMetric = {
  metricName: 'response_match_score',
  threshold: 0.7
};

The EvalConfigData interface defines the data required to initialize the evaluation configuration dialog.

Fields:

• evalMetrics: An array of EvalMetric objects representing currently selected metrics.
• metricsInfo: An array of MetricsInfo objects providing metadata (like valid ranges and intervals) for all available metrics.