testzeus-hercules 0.0.4

Hercules: The World's First Open-Source AI Agent for End-to-End Testing

💪 Hercules

Testing modern web applications can be difficult, with frequent changes and complex features making it hard to keep up. That’s where Hercules comes in. Hercules is the world’s first open-source testing agent, built to handle the toughest testing tasks so you don’t have to. It turns simple, easy-to-write Gherkin steps into fully automated end to end tests—no coding skills needed. Whether you’re working with tricky platforms like Salesforce or running tests in your CI/CD pipeline, Hercules adapts to your needs and takes care of the details. With Hercules, testing becomes simple, reliable, and efficient, helping teams everywhere deliver better software.

Using Hercules is as simple as feeding in your Gherkin features, and getting the results:

At TestZeus, we believe that trustworthy and open-source code is the backbone of innovation. That's why we've built Hercules to be transparent, reliable, and community-driven.

Our mission? To democratize and disrupt test automation, making top-tier testing accessible to everyone, not just the elite few. No more gatekeeping—everyone deserves a hero on their testing team!

⚡️ Features

Hercules is production ready, and packs a punch with features:

Gherkin In, Results Out

Hercules makes testing as simple as Gherkin in, results out. Just feed your end-to-end tests in Gherkin format, and watch Hercules spring into action. It takes care of the heavy lifting by running your tests automatically and presenting results in a neat JUnit format. No manual steps, no fuss—just efficient, seamless testing.

Free and Open Source

With Hercules, you're harnessing the power of open source with zero licensing fees. Feel free to dive into the code, contribute, or customize it to your heart's content. Hercules is as free as it is mighty, giving you the flexibility and control you need.

Salesforce Ready

Built to handle the most intricate UIs, Hercules conquers Salesforce and other complex platforms with ease. Whether it's complicated DOM or running your SOQL or Apex, Hercules is ready and configurable.

No Code Required

Say goodbye to complex scripts and elusive locators. Hercules is here to make your life easier with its no-code approach, taking care of the automation of Gherkin features so you can focus on what matters most—building quality software.

Multilingual

With multilingual support right out of the box, Hercules is ready to work with teams across the globe. Built to bridge language gaps, it empowers diverse teams to collaborate effortlessly on a unified testing platform.

Precisely Accurate

Hercules records video of the execution, and captures network logs as well, so that you dont have to deal with "It works on my computer".

No Maintenance

Autonomous and adaptive, Hercules takes care of itself with auto-healing capabilities. Forget about tedious maintenance—Hercules adjusts to changes and stays focused on achieving your testing goals.

UI Assertions

Grounded in the powerful foundations of TestZeus, Hercules tackles UI assertions with unwavering focus, ensuring that no assertion goes unchecked and no bug goes unnoticed. It's thorough, it's sharp, and it's ready for action.

CI/CD Ready

Run Hercules locally or integrate it seamlessly into your CI/CD pipeline. Docker-native and one-command ready, Hercules fits smoothly into your deployment workflows, keeping testing quick, consistent, and hassle-free.

With Hercules, testing is no longer just a step in the process—it's a powerful, streamlined experience that brings quality to the forefront.

---

⚙️ Installation and Usage

Hercules offers multiple ways to get started, catering to different user preferences and requirements.

Approach 1: Using PyPI Package

Installation

Install Hercules from PyPI:

pip install testzeus-hercules

Hercules uses Playwright to interact with web pages, so you need to install Playwright and its dependencies:

playwright install --with-deps

Basic Parameters

Once installed, you will need to provide some basic parameters to run Hercules:

• --input-file INPUT_FILE: Path to the input Gherkin feature file to be tested.
• --output-path OUTPUT_PATH: Path to the output directory. The path of JUnit XML result and HTML report for the test run.
• --test-data-path TEST_DATA_PATH: Path to the test data directory. The path where Hercules expects test data to be present; all test data used in feature testing should be present here.
• --project-base PROJECT_BASE: Path to the project base directory. This is an optional parameter; if you populate this, --input-file, --output-path, and --test-data-path are not required, and Hercules will assume all the three folders exist in the following format inside the project base:

PROJECT_BASE/
├── gherkin_files/
├── input/
│   └── test.feature
├── log_files/
├── output/
│   ├── test.feature_result.html
│   └── test.feature_result.xml
├── proofs/
│   └── User_opens_Google_homepage/
│       ├── network_logs.json
│       ├── screenshots/
│       └── videos/
└── test_data/
    └── test_data.txt

• --llm-model LLM_MODEL: Name of the LLM model to be used by the agent (recommended is gpt-4o, but it can take others).
• --llm-model-api-key LLM_MODEL_API_KEY: API key for the LLM model, something like sk-proj-k........

Running Hercules

After passing all the required parameters, the command to run Hercules should look like this:

testzeus-hercules --input-file opt/input/test.feature --output-path opt/output --test-data-path opt/test_data --llm-model gpt-4o --llm-model-api-key sk-proj-k.......

Supported AI Models for TestZeus-Hercules

• Anthropic Haiku: Compatible with Haiku 3.5 and above.
• Groq: Supports any version with function calling and coding capabilities.
• Mistral: Supports any version with function calling and coding capabilities.
• OpenAI: Fully compatible with GPT-4o and above. Note: OpenAI GPT-4o-mini is not supported.
• Ollama: Not supported based on current testing.

Execution Flow

Upon running the command:

• Hercules will start and attempt to open a web browser (default is Chromium).
• It will prepare a plan of execution based on the feature file steps provided.
• The plan internally expands the brief steps mentioned in the feature file into a more elaborated version.
• Hercules detects assertions in the feature file and plans the validation of expected results with the execution happening during the test run.
• All the steps, once elaborated, are passed to different tools based on the type of execution requirement of the step. For example, if a step wants to click on a button and capture the feedback, it will be passed to the click_using_selector tool.

Output and Logs

Once the execution is completed:

• Logs explaining the sequence of events are generated.
• The best place to start is the output-path, which will have the JUnit XML result file as well as an HTML report regarding the test case execution.
• You can also find proofs of execution such as video recordings, screenshots per event, and network logs in the proofs folder.
• To delve deeper and understand the chain of thoughts, refer to the chat_messages.json in the log_files. This will have exact steps that were planned by the agent.

Sample Feature File

Here's a sample feature file:

Feature: Account Creation in Salesforce

 Scenario: Successfully create a new account

   Given I am on the Salesforce login page
   When I enter my username "[e-mail]" and password "securePassword"
   And I click on the "Log In" button
   And I navigate to the "Accounts" tab
   And I click on the "New" button
   And I fill in the "Account Name" field with "Test Account"
   And I select the "Account Type" as "Customer"
   And I fill in the "Website" field with "www.testaccount.com"
   And I fill in the "Phone" field with "123-456-7890"
   And I click on the "Save" button
   Then I should see a confirmation message "Account Test Account created successfully"
   And I should see "Test Account" listed in the account records

Sample Result Screenshot

Sample HTML Result Screenshot

Sample XML Result Screenshot

---

Approach 2: Using Docker

For all the scale lovers, Hercules is also available as a Docker image.

Pull the Docker Image

docker pull testzeus/hercules:latest

Running Hercules in Docker

Run the container using:

docker run --env-file=.env \
  -v ./agents_llm_config.json:/testzeus-hercules/agents_llm_config.json \
  -v ./opt:/testzeus-hercules/opt \
  --rm -it testzeus/hercules:latest

• Environment Variables: All the required environment variables can be set by passing an .env file to the docker run command.
• LLM Configuration: If you plan to have complete control over Hercules and which LLM to use beyond the ones provided by OpenAI, you can pass agents_llm_config.json as a mount to the container. This is for advanced use cases and is not required for beginners. Refer to sample files .env-example and agents_llm_config-example.json for details and reference.
• Mounting Directories: Mount the opt folder to the Docker container so that all the inputs can be passed to Hercules running inside the container, and the output can be pulled out for further processing. The repository has a sample opt folder that can be mounted easily.
• Simplified Parameters: In the Docker case, there is no need for using --input-file, --output-path, --test-data-path, or --project-base as they are already handled by mounting the opt folder in the docker run command.

Browser Access in Docker

• While running in Docker mode, understand that Hercules has access only to a headless web browser.
• If you want Hercules to connect to a visible web browser, try the CDP URL option in the environment file. This option can help you connect Hercules running in your infrastructure to a remote browser like BrowserBase or your self-hosted grid.
• Use CDP_ENDPOINT_URL to set the CDP URL of the Chrome instance that has to be connected to the agent.

Output and Logs

After the command completion:

• The container terminates, and output is written in the mounted opt folder, in the same way as described in the directory structure.
• You will find the JUnit XML result file, HTML reports, proofs of execution, and logs in the respective folders.

---

Approach 3: Building and running from Source Code

For the hardcore enthusiasts, you can use Hercules via the source code to get a complete experience of customization and extending Hercules with more tools.

Prerequisites

• Ensure you have Python 3.11 installed on your system.

Steps to Run from Source

1. Clone the Repository

   git clone [e-mail]:test-zeus-ai/testzeus-hercules.git
   

2. Navigate to the Directory

   cd testzeus-hercules
   

3. Use Make Commands

   The repository provides handy make commands.

   • Use make help to check out possible options.

4. Install Poetry

   make setup-poetry
   

5. Install Dependencies

   make install
   

6. Run Hercules

   make run
   

   • This command reads the relevant feature files from the opt folder and executes them, putting the output in the same folder.

   • The opt folder has the following format:

     opt/
     ├── input/
     │   └── test.feature
     ├── output/
     │   ├── test.feature_result.html
     │   └── test.feature_result.xml
     ├── log_files/
     ├── proofs/
     │   └── User_opens_Google_homepage/
     │       ├── network_logs.json
     │       ├── screenshots/
     │       └── videos/
     └── test_data/
         └── test_data.txt
     

7. Interactive Mode

   You can also run Hercules in interactive mode as an instruction execution agent, which is more useful for RPA and debugging test cases and Hercules's behavior on new environments while building new tooling and extending the agents.

   make run-interactive
   

   • This will trigger an input prompt where you can chat with Hercules, and it will perform actions based on your commands.

---

📝 Configuration Details

Understanding the Environment File (.env)

To configure Hercules in detail:

• Copy the base environment file:

  cp .env-example .env
  

• Hercules is capable of running in two configuration forms:

  1. Using single LLM for all work

     • For all the activities within the agent, initialize LLM_MODEL_NAME and LLM_MODEL_API_KEY.
     • If using a non-OpenAI hosted solution but still OpenAI LLMs (something like OpenAI via Groq), then pass the LLM_MODEL_BASE_URL URL as well.

  2. Custom LLMs for different work or using hosted LLMs

     • If you plan to configure local LLMs or non-OpenAI LLMs, use the other parameters like AGENTS_LLM_CONFIG_FILE and AGENTS_LLM_CONFIG_FILE_REF_KEY.
     • These are powerful options and can affect the quality of Hercules outputs.

• Hercules considers a base folder that is by default ./opt but can be changed by the environment variable PROJECT_SOURCE_ROOT.

• Connecting to an Existing Chrome Instance

  • This is extremely useful when you are running Hercules in Docker for scale.
  • You can connect Hercules running in your infrastructure to a remote browser like BrowserBase or your self-hosted grid.
  • Use CDP_ENDPOINT_URL to set the CDP URL of the Chrome instance that has to be connected to the agent.

• Controlling Other Behaviors

  You can control other behaviors of Hercules based on the following environment variables:

  • HEADLESS=true
  • RECORD_VIDEO=false
  • TAKE_SCREENSHOTS=false
  • BROWSER_TYPE=chromium (options: firefox, chromium)
  • CAPTURE_NETWORK=false

Understanding agents_llm_config-example.json

• It's a list of configurations of LLMs that you want to provide to the agent.

• Example:

  {
    "mistral-large-agente": {
      "planner_agent": {
        "model_name": "mistral",
        "model_api_key": "",
        "model_base_url": "https://...",
        "system_prompt": "You are a web automation task planner....",
        "llm_config_params": {
          "cache_seed": null,
          "temperature": 0.1,
          "top_p": 0.1
        }
      },
      "browser_nav_agent": {
        "model_name": "mistral",
        "model_api_key": "",
        "model_base_url": "https://...",
        "system_prompt": "You will perform web navigation tasks with the functions that you have...\nOnce a task is completed, confirm completion with ##TERMINATE TASK##.",
        "llm_config_params": {
          "cache_seed": null,
          "temperature": 0.1,
          "top_p": 0.1
        }
      }
    }
  }
  

• The key is the name of the spec that is passed in AGENTS_LLM_CONFIG_FILE_REF_KEY, whereas the Hercules information is passed in sub-dicts planner_agent and browser_nav_agent.

• Note: This option should be ignored until you are sure what you are doing. Discuss with us while playing around with these options in our Discord communication.

---

🦾 Architecture

Multi-Agentic Solution

Hercules leverages a multi-agent architecture based on the AutoGen framework. Building on the foundation provided by the AutoGen framework, Hercules's architecture leverages the interplay between tools and agents. Each tool embodies an atomic action, a fundamental building block that, when executed, returns a natural language description of its outcome. This granularity allows Hercules to flexibly assemble these tools to tackle complex web automation workflows.

System View

The diagram above shows the configuration chosen on top of AutoGen architecture. The tools can be partitioned differently, but this is the one that we chose for the time being. We chose to use tools that map to what humans learn about the web browser rather than allow the LLM to write code as it pleases. We see the use of configured tools to be safer and more predictable in its outcomes. Certainly, it can click on the wrong things, but at least it is not going to execute malicious unknown code.

Agents

At the moment, there are two agents:

1. Planner Agent: Executes the planning and decomposition of tasks.
2. Browser Navigation Agent: Embodies all the tools for interacting with the web browser.

Tools Library

At the core of Hercules's capabilities is the Tools Library, a repository of well-defined actions that Hercules can perform; for now, web actions. These tools are grouped into two main categories:

• Sensing Tools: Tools like get_dom_with_content_type and geturl that help Hercules understand the current state of the webpage or the browser.
• Action Tools: Tools that allow Hercules to interact with and manipulate the web environment, such as click, enter_text, and openurl.

Each tool is created with the intention to be as conversational as possible, making the interactions with LLMs more intuitive and error-tolerant. For instance, rather than simply returning a boolean value, a tool might explain in natural language what happened during its execution, enabling the LLM to better understand the context and correct course if necessary.

Implemented Tools

• Sensing Tools

  • geturl: Fetches and returns the current URL.
  • get_dom_with_content_type: Retrieves the HTML DOM of the active page based on the specified content type.
    • text_only: Extracts the inner text of the HTML DOM. Responds with text output.
    • input_fields: Extracts the interactive elements in the DOM (button, input, textarea, etc.) and responds with a compact JSON object.
    • all_fields: Extracts all the fields in the DOM and responds with a compact JSON object.
  • get_user_input: Provides the orchestrator with a mechanism to receive user feedback to disambiguate or seek clarity on fulfilling their request.

• Action Tools

  • click: Given a DOM query selector, this will click on it.
  • enter_text: Enters text in a field specified by the provided DOM query selector.
  • enter_text_and_click: Optimized method that combines enter_text and click tools.
  • bulk_enter_text: Optimized method that wraps enter_text method so that multiple text entries can be performed in one shot.
  • openurl: Opens the given URL in the current or new tab.

DOM Distillation

Hercules's approach to managing the vast landscape of HTML DOM is methodical and essential for efficiency. We've introduced DOM Distillation to pare down the DOM to just the elements pertinent to the user's task.

In practice, this means taking the expansive DOM and delivering a more digestible JSON snapshot. This isn't about just reducing size; it's about honing in on relevance, serving the LLMs only what's necessary to fulfill a request. So far, we have three content types:

1. Text Only: For when the mission is information retrieval, and the text is the target. No distractions.
2. Input Fields: Zeroing in on elements that call for user interaction. It's about streamlining actions.
3. All Content: The full scope of distilled DOM, encompassing all elements when the task demands a comprehensive understanding.

It's a surgical procedure, carefully removing extraneous information while preserving the structure and content needed for the agent's operation. Of course, with any distillation, there could be casualties, but the idea is to refine this over time to limit/eliminate them.

Since we can't rely on all web page authors to use best practices, such as adding unique IDs to each HTML element, we had to inject our own attribute (mmid) in every DOM element. We can then guide the LLM to rely on using mmid in the generated DOM queries.

To cut down on some of the DOM noise, we use the DOM Accessibility Tree rather than the regular HTML DOM. The accessibility tree, by nature, is geared towards helping screen readers, which is closer to the mission of web automation than plain old HTML DOM.

The distillation process is a work in progress. We look to refine this process and condense the DOM further, aiming to make interactions faster, cost-effective, and more accurate.

---

🔬 Testing and Evaluation: QEvals

We wanted to ensure that Hercules stands up to the task of end-to-end testing with immense precision. So, we have run Hercules through a wide range of tests such as running APIs, interacting with complex UI scenarios, clicking through calendars, or iframes. A full list of evaluations can be found in the tests folder.

Running Tests

To run the full test suite, use the following command:

make test

To run a specific test:

make test-case

Hercules builds on the work done by WebArena and Agent-E, and beyond that, to iron out the issues in the previous, we have written our own test cases catering to complex QA scenarios and have created tests in the ./tests folder.

---

💡 Opinions

We believe that great quality comes from opinions about a product. So we have incorporated a few of our opinions into the product design. We welcome the community to question them, use them, or build on top of them. Here are some examples:

1. Gherkin is a Good Enough Format for Agents: Gherkin provides a semi-structured format for the LLMs/AI Agents to follow test intent and user instructions. It provides the right amount of grammar (verbs like Given, When, Then) for humans to frame a scenario and agents to follow the instructions.

2. Tests Should Be Atomic in Nature: Software tests should be atomic because it ensures that each test is focused, independent, and reliable. Atomic tests target one specific behavior or functionality, making it easier to pinpoint the root cause of failures without ambiguity.

   Here's a good example (Atomic Test):

   Feature: User Login
   
   Scenario: Successful login with valid credentials
   
     Given the user is on the login page
     When the user enters valid credentials
     And the user clicks the login button
     Then the user should see the dashboard
   

   A non-atomic test confuses both the tester and the AI agent.

3. Open Core and Open Source: Hercules is built on an open-core model, combining the spirit of open source with the support and expertise of a commercial company, TestZeus. By providing Hercules as open source (licensed under AGPL v3), TestZeus is committed to empowering the testing community with a robust, adaptable tool that's freely accessible and modifiable. Open source offers transparency, trust, and collaborative potential, allowing anyone to dive into the code, contribute, and shape the project's direction.

---

🪙 Token Usage

Hercules is an AI-native solution and relies on LLMs to perform reasoning and actions. Based on our experiments, we have found that a complex use case as below could cost up to $0.20 using OpenAI's APIs gpt-4o, check the properties printed in testcase output to calculate for your testcase:

Feature: Account Creation in Salesforce

 Scenario: Successfully create a new account

   Given I am on the Salesforce login page
   When I enter my username "[e-mail]" and password "securePassword"
   And I click on the "Log In" button
   And I navigate to the "Accounts" tab
   And I click on the "New" button
   And I fill in the "Account Name" field with "Test Account"
   And I select the "Account Type" as "Customer"
   And I fill in the "Website" field with "www.testaccount.com"
   And I fill in the "Phone" field with "123-456-7890"
   And I click on the "Save" button
   Then I should see a confirmation message "Account Test Account created successfully"
   And I should see "Test Account" listed in the account records

---

⭐️ Difference from Other Tools

Hercules isn't just another testing tool—it's an agent. Powered by synthetic intelligence that can think, reason, and react based on requirements, Hercules goes beyond simple automation scripts. We bring an industry-first approach to open-source agents for software testing. This means faster, smarter, and more resilient testing cycles, especially for complex platforms.

With industry-leading performance and a fully open-source foundation, Hercules combines powerful capabilities with community-driven flexibility, making top-tier testing accessible and transformative for everyone.

---

🚀 High-Level Roadmap

• Enhanced LLM Support: Integration with more LLMs and support for local LLM deployments.
• Advanced Tooling: Addition of more tools to handle complex testing scenarios and environments.
• Improved DOM Distillation: Refinements to the DOM distillation process for better efficiency and accuracy.
• Community Contributions: Encourage and integrate community-driven features and tools.
• Extensive Documentation: Expand documentation for better onboarding and usage.
• Bounty Program: Launch a bounty program to incentivize contributions.

---

🤗 Contribution

We welcome contributions from the community!

• Read the CONTRIBUTING.md file to get started.
• Bounty Program: Stay tuned for upcoming opportunities! 😀

How to Contribute Back

1. Developing Tools

   • If you are developing tools for Hercules and want to contribute to the community, make sure you place the new tools in the additional_tools folder in your Pull Request.

2. Fixes and Enhancements

   • If you have a fix on sensing tools that are fundamental to the system or something in prompts or something in the DOM distillation code, then put the changes in the relevant file and share the Pull Request.

Extending and Attaching Tools

1. Creating a New Tool

   • You can start extending by adding tools to Hercules.
   • Refer to testzeus_hercules/core/tools/sql_calls.py as an example of how to create a new tool.
   • The key is the decorator @tool over the method that you want Hercules to execute.
   • The tool decorator should have a very clear description and name so that Hercules knows how to use the tool.
   • Also, in the method, you should be clear with annotations on what parameter is used for what purpose so that function calling in the LLM works best.

2. Adding the Tool

   • Once you have created the new tools files in some folder path, you can pass the folder path to Hercules in the environment variable so that Hercules can read the new tools during the boot time and make sure that they are available during the execution.
   • Use ADDITIONAL_TOOL_DIRS to pass the path of the new tools folder where you have kept the new files.

3. Direct Addition (Not Recommended)

   • In case you opt for adding the tools directly, then just put your new tools in the testzeus_hercules/core/tools path of the cloned repository.

   • Then make sure you import your tool module in the testzeus_hercules/core/agents/browser_nav_agent.py file as:

     from testzeus_hercules.core.tools.sql_calls import *
     

   • Note: This way is not recommended. We prefer you try to use the ADDITIONAL_TOOL_DIRS approach.

---

🤙 Contact Us

Join us at our Discord server to connect with the community, ask questions, and contribute.

---

✍️ Examples

• Salesforce Examples: Link
• Wrangler Example: Link

---

🙇 Credits

Hercules would not have been possible without the great work from the following sources:

1. Agent-E
2. Q*
3. Agent Q
4. Autogen

The Hercules project is inferred and enhanced over the existing project of Agent-E. We have improved lots of cases to make it capable of doing testing, especially in the area of complex DOM navigation and iframes. We have also added new tools and abilities (like Salesforce navigation) to Hercules so that it can perform better work over the base framework we had picked.

Hercules also picks some inspiration from the legacy TestZeus repo here.

---

With Hercules, testing is no longer just a step in the process—it's a powerful, streamlined experience that brings quality to the forefront.