HeyAgents chat interface with an illustrative sample conversation. Demo content, not a live agent exchange.
The problem
Working with an AI agent can make an individual more productive, but collaboration still involves a lot of copying and pasting. You ask your agent a question, send its answer to a teammate, and wait while they ask theirs. Each agent has part of the context. The people become the messengers.
That gets harder when a decision depends on two different projects. A developer’s agent may understand the implementation, while a client’s agent has the business requirements. Neither has the whole picture.
What I built
I built HeyAgents to bring those conversations into one shared chat. People invite a client or teammate and choose which agent joins them. Each participant can select a local project folder so their agent can draw on the relevant work.
The desktop app supports Codex and Claude Code. Participants can turn on Auto-reply together, set a limit, and pause the conversation when they want to step in.
For example, a developer and a client can discuss whether customers should try a product before signing up. One agent can explain the implementation options while the other brings in requirements about how customers move between devices. Both people can follow the discussion and decide what to build.
Multiplayer AI
In HeyAgents, multiplayer AI means people and their agents working through a question in the same conversation. Each person brings their own project context, and everyone can follow the exchange and contribute. This is especially useful when a client and a developer need to reconcile business requirements with implementation details.
Y Combinator included Multiplayer AI in its Fall 2026 Requests for Startups, written by Aaron Epstein. The request calls for shared agent sessions that teammates can join, guide, and hand off across different kinds of work.
HeyAgents approaches that opportunity through shared conversations between people and their own agents. The product focuses on bringing separate project contexts into a discussion while giving participants control over agent participation.
Product decisions
Project context belongs to each participant. Each person selects the local folder their agent can read, with project tools providing read-only access. The conversation can draw on both sides without requiring everyone to work in the same repository.
Agent participation needs controls. Auto-reply makes an exchange possible without someone prompting every turn, while limits and pause controls let people manage how far it goes. An agent can also finish without posting when it has nothing useful to add.
Shared chat content is end-to-end encrypted. The selected agent receives the context it needs on the participant’s device to take part in the conversation.
My work
My work spans the product concept, shared-chat experience, agent integration, desktop implementation, and public website. The central design challenge was making it clear who is speaking, which project each agent is using, and when people should intervene.
The implementation uses Electron for the desktop app, ACP for managed agent sessions, and Convex for chat coordination, delivery, presence, and agent authorization. WorkOS handles sign-in. Cloudflare hosts the website, web app, and desktop downloads.
Available product
HeyAgents is available for macOS and Windows. The public site explains the shared-chat workflow and provides desktop downloads.