Gstack Agents: Open-Source AI Expert Panel Joins Your Google Meet Video Conferences

When AI Role-Playing Enters the Real Meeting Room

Imagine this scenario: you're demoing your product in Google Meet, but on the other end of the screen, instead of colleagues, there's a group of AI experts sporting 3D avatars and delivering critiques—one playing the CEO, another the Chief Strategy Officer, and yet another simulating a YC partner's "office hours" experience. They take turns speaking, critique your shared screen content in real time, and drop key points into the chat. This is exactly the scenario that Gstack Agents, an open-source project recently featured on Product Hunt, aims to create.

The project was built by Anand Balakrishnan, inspired by the "gstack specialists" persona library open-sourced by YC head Garry Tan. It has already received 118 upvotes on Product Hunt, ranking 11th on the daily leaderboard, and is categorized under open-source tools, developer tools, and artificial intelligence.

Some background on YC Office Hours is worth elaborating: Y Combinator is one of the world's most influential startup accelerators, having incubated well-known companies like Airbnb, Stripe, and Dropbox. One of its core operating mechanisms is "Office Hours"—regular one-on-one meetings between YC partners and portfolio founders, where partners deliver extremely direct, high-density interrogative feedback on product direction, business models, and growth strategies. This format is regarded by Silicon Valley entrepreneurs as an incredibly valuable "mental stress test." Garry Tan, as YC's current CEO, open-sourced his persona library in an attempt to digitally replicate this high-quality feedback model.

What Is Gstack Agents

An AI Expert Panel That Can "Join" Your Meeting

The core value proposition of Gstack Agents is pulling AI out of chat windows and into real video conferences. The entire expert roster includes 18+ AI personas—CEO, CSO (Chief Strategy Officer), QA Lead, a partner simulating YC office hours, and more. These personas join your Google Meet as voice bots, equipped with 3D virtual avatars.

Their behavioral logic is designed to be quite realistic: they introduce themselves in-persona, take turns speaking, verbally critique the screen content you're sharing, and leave written notes in the meeting chat. In other words, it simulates a real "multi-persona review meeting" rather than a single AI assistant's Q&A session.

Free, MIT Open-Source, with You Controlling the Inference Core

This project has several noteworthy technical positioning points. First, it's completely free and open-sourced under the MIT license. The MIT license is one of the most permissive licenses in open-source software, allowing anyone to freely use, copy, modify, merge, publish, distribute, sublicense, and even sell copies of the software—the only requirement being that copyright notices are included in all copies. Compared to "copyleft" licenses like GPL (which require derivative works to also be open-sourced), the MIT license imposes virtually no restrictions on commercial use. This means companies can build closed-source commercial products on top of Gstack Agents without disclosing their code, significantly reducing legal risk for commercialization and making it easier to attract enterprise developers to the ecosystem.

Second, the team specifically emphasizes that "the brain is your own coding-agent session"—meaning the inference core driving these AI personas isn't locked into any closed service but connects to the user's own coding agent session. Specifically, a "coding-agent session" refers to an AI agent runtime environment centered on code generation and execution capabilities, such as the persistent conversational contexts provided by tools like Cursor, Windsurf, or Claude Code. Under this architecture, Gstack itself only handles the "presentation layer" work—persona orchestration, speech synthesis, meeting integration—while the actual cognitive reasoning is handled by external models. This "thin client" design both avoids the project team bearing expensive GPU inference costs and allows users to flexibly choose GPT-4, Claude, Gemini, or other models to power the personas.

The entire system is built on AgentCall, the underlying framework that supports multiple voice agents collaboratively entering real-time meetings. The core technical challenges AgentCall must solve include: multi-agent turn-taking—ensuring multiple AIs don't speak simultaneously or remain silent for too long; low-latency pipelines for real-time text-to-speech (TTS) and speech-to-text (STT); and meeting protocol adaptation—integrating AI agents as compliant participants into WebRTC audio/video streams on platforms like Google Meet. Traditional voice AI applications are mostly single-agent, while multi-agent scenarios require solving additional complexities like speaking priority, context sharing, and conversation state synchronization, making AgentCall a critical "orchestration layer" in the tech stack.

Why This Approach Deserves Attention

From "Conversational AI" to "Present AI"

Over the past two years, the interaction paradigm of the vast majority of AI applications has remained stuck in "you ask, I answer" dialog boxes. Gstack Agents represents a different direction: having AI exist as a "meeting participant" within humans' existing collaboration scenarios. The difference in this sense of "presence" isn't just a formal innovation—when AI can observe your shared screen, chime in at appropriate moments, and express opinions from a specific role's standpoint, it more closely resembles a real review consultant rather than a tool.

From a broader perspective, the evolution of AI interaction paradigms has roughly gone through three stages: command-line style (users input precise commands to get results), conversational (multi-turn interaction in natural language, like ChatGPT), and the emerging "present" or "ambient" style (Ambient AI). The core characteristic of present AI is that it no longer waits for users to initiate conversations but continuously perceives environmental context (such as screen content, meeting progress, code changes) and proactively intervenes at appropriate moments. The theoretical foundation of this paradigm can be traced back to Mark Weiser's 1991 vision of "Ubiquitous Computing"—technology should blend into the environment rather than requiring users to deliberately interact with it. Gstack Agents concretizes this concept in the form of "AI meeting participants," representing an identity leap from tool to colleague.

For entrepreneurs, this design is particularly appealing. Scenarios like YC office hours, simulated investor interrogations, and product reviews all fundamentally require "multiple different perspectives present simultaneously." Using AI personas to batch-replicate these perspectives costs far less than actually meeting with multiple experts.

The Multi-Perspective Review Value of Role Division

Advice from a single large model often tends toward averaged, moderate answers. By having CEO, CSO, QA Lead, and other personas speak separately, Gstack effectively uses prompt engineering and persona settings to force models to output opinions from different stakeholder positions.

Here it's important to understand the technical principles of prompt engineering in role-playing: through carefully designed "System Prompts," each persona can be injected with personality settings, professional background descriptions, and behavioral constraint rules. For example, a "CEO persona's" system prompt might include instructions like: focus on revenue models and market size, ask questions from a decision-maker's perspective, prefer concise and direct communication style. Research shows that persona-based prompting can effectively activate domain-specific knowledge patterns learned from pre-training data, producing more focused and position-driven outputs than generic prompts.

The CEO might focus on the business model, the CSO scrutinizes strategic moats, and the QA Lead zeros in on product defects. This structured multi-perspective review often has more practical value than a single vague piece of feedback. However, it's worth noting that persona consistency remains a known weakness of current large language models—phenomena like personas "drifting" or confusing different role positions during long conversations are not uncommon, posing a potential challenge to Gstack's actual experience quality.

Real-World Opportunities and Concerns

Engineering Challenges of Voice Coordination and Real-Time Performance

Stuffing multiple AI voice agents into a real-time video conference is technically no easy feat. Turn-taking scheduling, voice naturalness, 3D avatar rendering, real-time understanding of shared screen content—every link tests the engineering implementation. The real-time screen understanding component in particular requires continuous parsing of screen frames from the video stream through multimodal models (like GPT-4V or Gemini Pro Vision), with parsing results injected into each persona's conversational context in real time, placing extremely high demands on both latency and computational resources.

The project relies on the AgentCall framework to handle these complexities, but judging from the early-stage heat of only 10 comments on Product Hunt, it currently looks more like a proof of concept and developer toy, still some distance from a production-grade stable experience.

Feedback Quality Depends on the Model You Connect

A point worth repeatedly emphasizing: since Gstack delegates the inference core to the user's own coding-agent session, the quality of AI feedback is highly dependent on the capability of the model the user connects to, as well as the design precision of the persona prompts. This is both where its flexibility lies—you can drive it with any powerful model—and its source of uncertainty. When role-playing is done well, it's a high-quality review; done poorly, it may devolve into several AI avatars talking to themselves.

From a practical standpoint, connecting different levels of models produces significantly different experiences: using a GPT-4-level model might produce high-quality strategic insights, while using smaller open-source models might significantly compromise persona maintenance and professional depth. Additionally, when multiple personas share the same model backend, how to ensure they truly produce differentiated rather than homogenized viewpoints is also a prompt engineering challenge requiring careful tuning.

Conclusion

The significance of Gstack Agents may not lie in how polished it is at this stage, but in how it demonstrates a new form of AI application: embedding AI into humans' real collaborative workflows in a persona-driven, voice-enabled, present manner. For those of us accustomed to chat boxes, this is a direction worth watching. And its positioning as MIT open-source, free, and connectable to your own models lowers the barrier for developers to experiment. If you're looking for a more immersive AI product review experience, this project is at least worth cloning and running once.