Gstack puts AI agents in your meetings as talking specialists

Gstack ships voice agents that join Google Meet as speaking participants. The agents take on specialist roles: CEO, CSO, QA lead, a YC office-hours partner, and 15 others. They open in character, take turns, critique a shared screen out loud, and drop notes in chat. Garry Tan open-sourced the collection under MIT on July 27, according to the announcement on Product Hunt.

The product is built on AgentCall, a voice-agent infrastructure layer. Each specialist runs as a coding-agent session, meaning the brain is a language model prompted with a role and a knowledge base. The agents have 3D avatars. They appear in the meeting roster as real participants. They speak in natural turns. They can handle overlapping speech by prioritizing the most recent speaker, according to the gstack documentation.

The immediate use case is Y Combinator interview prep. Founders can bring a YC-office-hours partner agent into a practice pitch. The agent asks questions, pushes back on weak spots, and scores the presentation. That is a narrow application. The broader pattern is more interesting: a company open-sourcing a library of specialist voice agents that any team can deploy into any meeting.

Tan is Y Combinator’s CEO. He is also an angel investor and a former partner at the fund. That he personally open-sourced this collection signals something about where YC sees the market going. YC funds roughly 500 companies per batch. Many of those companies are building AI agents. Many of those agents are text-only. Voice agents that participate in real-time conversation are the next layer. Gstack is a reference implementation of that layer.

The technical architecture matters. Each agent is a coding-agent session, not a fine-tuned model. That means the specialist behavior is entirely prompt-driven. The same model can be a CEO in one meeting and a QA lead in the next, depending on the system prompt and the knowledge base uploaded. The agent does not need a separate login. It joins via a meeting link or dial-in number, the same way a human guest would. It does not independently record audio unless explicitly enabled, according to the FAQ.

The limitations are visible. Voice synthesis can still sound artificial in some contexts. The agent may struggle with complex or ambiguous questions. Overlapping speech handling works but is not perfect. The agent cannot read body language or tone shifts that a human would catch. These are not bugs in gstack specifically. They are limits of current voice AI.

What gstack does well is remove the scheduling bottleneck. A team that needs a QA specialist for a 30-minute review no longer needs to find a human QA engineer with availability. They spin up the agent. The agent brings the knowledge base. The agent speaks in real time. The team gets feedback without the calendar negotiation.

The open-source decision is the most consequential part of the announcement. MIT license means any company can fork the agent library, customize the roles, and deploy into their own meeting infrastructure. That accelerates the spread of voice agents in professional settings. It also lowers the barrier for competitors. A startup building meeting agents now has a reference architecture to study and improve on.

The market for meeting agents is already crowded. Otter.ai transcribes. Fathom summarizes. Fireflies notetakes. None of those products speak. None of them join as a participant with a role and a voice. Gstack is in a different category: the agent as a peer, not a scribe.

The question is whether teams actually want an AI voice in the room. The first reaction from many users will be discomfort. An AI that speaks, interrupts, and critiques is different from an AI that silently transcribes. The social dynamics of a meeting change when a non-human participant has a voice and an avatar. Some teams will reject it. Others will find it liberating.

The YC interview use case is a smart beachhead. Practice pitches are low stakes. The feedback is high value. The agent does not need to be perfect. It just needs to be better than no practice partner. For a founder preparing for a 10-minute YC interview, a voice agent that asks tough questions and scores the answers is a clear win.

The long-term implication is that every meeting will eventually have multiple AI participants. Not just one scribe, but a team of specialists. A legal reviewer. A security auditor. A product critic. Each with its own knowledge base and personality. Each joining and leaving as needed. Gstack is a prototype of that future.

The open-source release means the future arrives faster. Any developer can clone the repo, add a new specialist, and deploy it into their team’s meetings. The network effect is real: the more specialists the community builds, the more valuable the platform becomes. Tan is betting that the value of the ecosystem exceeds the value of a closed product.

The bet is not guaranteed. Voice agents still have a quality ceiling. The artificial tone, the lag in response, the occasional misinterpretation. Those problems will improve with better models and lower latency. They will not disappear overnight. But the trajectory is clear. Voice agents in meetings are moving from demo to deployment.

What to watch next: whether the open-source community adds specialists that gstack did not ship. A compliance officer. A technical architect. A customer support escalation agent. Each new specialist tests the limits of prompt-driven voice agents. Each one expands the surface area of what an AI meeting participant can do.

The most concrete observation from the gstack launch is this: the agents speak, they have avatars, and they are free to fork. The rest is an experiment in whether teams want AI peers in their meetings.