Google's Opal: A New No-Code AI Workflow Builder

It has been several years since ChatGPT first introduced the chat interface for LLMs. Whether we're looking at OpenAI's ChatGPT, Gemini, or Claude, these interfaces have evolved massively, incorporating various buttons for different functionalities. Concurrently, the use of APIs has progressed from simple frameworks like LangChain to more advanced agentic features. A key development in this space is the blurring of lines between programming and simple text input.

The Rise of No-Code LLM Workflow Tools

A major champion in this domain has been n8n, a tool many have adopted to build LLM application workflows without extensive coding knowledge. Other platforms like Lindy have also been available for some time, and numerous similar solutions are emerging.

The latest entrant into this competitive landscape is Google, specifically through Google Labs. Google Labs is a product-focused team, widely known for NotebookLM, which is now a couple of years old. NotebookLM began as an internal project to create a RAG system that simplifies document and web link integration for conversational AI. Its popularity surged after the introduction of a voice and podcasting feature, which was originally a separate project called GenFM before being merged into NotebookLM.

Introducing Opal: Google Labs' Latest Preview

This brings us to the newest public preview from Google Labs: Opal. Like other Google Labs products, Opal is released as a preview for public testing and feedback. Currently, it is available only in the US, though access may be possible with a VPN.

Opal aims to provide a simplified, no-code experience for building LLM workflows, similar to n8n or Lindy. While not yet at the level of hardcore agent development, it allows users to chain prompts, utilize various models, and access Google's tools. This enables the prototyping of workflows for later coding or the creation of mini-apps for immediate use.

The concept is straightforward: you describe the tool you want to build, or you can remix existing examples. Opal then maps out a series of steps. Some steps might involve tools for web searches or fetching webpage content, while others use Gemini models for tasks like outlining, writing blog posts, or creating images. The core idea is to offer a no-code method for creating, saving, and reusing these workflows.

A Practical Walkthrough of Opal

Opal features a gallery with several pre-built applications. Let's explore by remixing one of their existing apps.

We'll start with the "blog post generator." To run it, you provide a topic. For this example, the topic is:

"A blog post about Opal, Google Lab's latest product."

Once the process starts, Opal moves through a sequence of steps: searching the web, retrieving web pages, writing an outline, drafting the post, and creating a banner image. These tasks can run concurrently. The final output is a displayed blog post.

By looking at the console, you can inspect what each step is doing. For instance, the research step might call Gemini 1.5 Flash and perform a web search. This provides valuable insight into how you could structure your own agents and workflows.

The initial output might be a bit generic. For example, a generated post titled "Unlocking Your Inner App Developer: A Deep Dive into Google's New AI Gem, Opal" might be factually correct but lack a specific tone.

Customizing and Enhancing Workflows

This is where customization comes in. You can edit each step of the workflow. For instance, you might notice the writing step uses Gemini 1.5 Flash while the research used a different version. You can change the model for any step to Gemini 1.5 Pro, Veo, or Lyria.

For image generation, the default might be an older model. Let's try changing it to Imagen 3 and modify the prompt to include the topic in the banner text. To do this, you need to pass the topic as an input to the image generation step.

You can also add new user inputs. For example, let's add an input for the reader's persona:

"Please provide the persona of the intended reader."

This new input can be wired into the research and writing steps. The prompt for the writing step can be updated to:

"Produce a blog post using the research report for the reader persona."

After resetting and running the workflow again with the new inputs, for example, an IT worker looking for automation solutions, the process will incorporate this new context. The research will create a plan tailored to this persona, and the final blog post will be heavily customized. The banner image generated by Imagen 3 might even include text like "Google Opal," showcasing its advanced capabilities.

Building from Scratch

What if you want to create a tool from the ground up? You can simply describe what you want to build. For example:

"I wanna make a tool that does literature reviews on arXiv papers and is able to find related arXiv papers to an arXiv paper that I upload."

Opal will then generate a workflow based on your description. It might create inputs for an arXiv paper URL and a literature review topic, a step to fetch the URL, and another to search for related papers. You can then add new nodes, such as a "deep research" node using Gemini 1.5 Flash, to extract author information or perform other tasks.

The platform integrates various Google tools, including Google Search and different image and audio generation models like AudioLM, allowing you to build a wide range of applications, from research assistants to podcast generators.

The Future of No-Code AI

Opal is a promising new product from Google Labs. As a preview, it represents just the first version, and we can expect more features to be added over time as user feedback comes in, much like the evolution of NotebookLM.

For those who want to create LLM-powered apps and tools without diving deep into code, platforms like Opal open up a new world of possibilities. They allow for the rapid creation of generative AI workflows for numerous tasks. Furthermore, the prompts and logic developed in Opal can be extracted and implemented in code for more complex, production-level applications.