Hive — Multiplayer AI coding for engineering and Product

Two engineers. Two AI agents. One shared project. Hive lets Claude and Gemini collaborate in real time so your whole team sees the decisions, not just the final PR.

Comment

Inspiration

PRs show what. Never why. And when two engineers use AI agents today, all that reasoning is invisible to everyone else on the team. We wanted to fix that.

What it does

Hive is the first multiplayer AI coding environment. It connects Claude Code and Gemini CLI through a shared relay — two agents, one project, in real time.

Engineers see each other's agent decisions as they happen. PMs get a chat interface that answers "why did we build it this way?" with the source session cited. Drop a whiteboard screenshot in and ask if the team actually built it that way — Gemini Vision cross-references the diagram against every agent session.

How we built it

Claude Code and Gemini CLI both write sessions to disk as JSONL files. We watch those files, stream events to a shared relay on DigitalOcean, and normalize Claude + Gemini output into a shared schema.

Railtracks orchestrates all AI inference — summarization, feature bucketing, PM Q&A, vision queries. Gemini Flash buckets agent events into readable feature stories. Claude Sonnet handles nuanced PM questions. Assistant UI powers the chat.

Challenges we ran into

Feature bucketing is an editorial problem, not a data problem. A timeline of 200 agent events is as useless as no timeline. The AI has to act like an editor — group by meaning, surface open questions first.

The other challenge: building without reading the code. When you're moving fast and delegating entirely to AI agents, bugs appear not because things are underspecified — but because the agent makes engineering decisions you never see. Iteration is the hard part, not the initial build.

Accomplishments that we're proud of

Claude is wired up and live. The relay is running. Two AI agents from two different companies sharing context through one relay — we don't know of another tool that does this.

What we learned

AI-native design means the interface comes to you. You don't go to it. And the simplest architecture — just watch the files — is usually the right one.

What's next for Hive

Git integration to link agent sessions to PRs automatically. More agent adapters. Meeting → session pipeline so every agent starts a sprint with the full context of what was decided.

The vision: a world where the PR is the least interesting artifact a team produces.

Built With

Share this project:

Updates