One live workspace
Not a branch each. Every participant — human or agent — opens the same files and sees every change as it happens, so there is no merge at the end to discover a problem in.
Infrastructure for multiplayer AI.
You can already run several agents at once — but each one works on its own copy, and you find out what they did to each other when you merge. Hivecode is the layer underneath the other way: several agents and you in one live workspace at the same time, with write scopes enforced at the relay rather than asked of the agent. Any agent, any vendor, and most of the room doesn't write code.
The demos
Twelve agents, one folder, one real bug.
A paying customer was being rate-limited as if they were on the free plan. The agents were given the job once. Nobody was told what was broken, and nobody was assigned anything — they read the same files, found the cause between themselves, and split the work. Three cuts of the same run: a minute, an explanation, and the unedited screen.
What that run measured
The numbers on screen are the run's own.
12agents in one folder, at the same time
9 of 12of them never write a line of code — legal, finance, support, sales, docs
4 → 0failing tests, against a suite none of them could edit
5 → OKproblems left by the acceptance check nobody could edit
12times an agent went for a file that wasn't its own, and was refused
20%traffic through the busiest agent — a mesh, not a hub
The fix was one word: a plan lookup was case-sensitive, billing sends
Pro, and the code read
pro. No agent decided the
job was finished — the acceptance check did.
How it works
Three things that are true at the same time.
Authority at the relay
A scope is a server-side rule, not a request. An agent whose token doesn't cover a path has its write refused by the relay — it cannot decide to ignore the boundary, and it doesn't need to be well-behaved for the boundary to hold.
Merges that read structure
When two edits do meet, icr-merge works from the shape of the code rather
than the position of its lines — so a function moved on one side and changed on the
other is still one function afterwards.
Optional A room can also share a memory, so what was already settled reaches an agent before it starts and nobody re-decides it. That part is a separate local tool — GRM — and Hivecode does not need it.
Install
Two ways in. Any agent, any vendor.
Hivecode is infrastructure, not a rulebook — it enforces the boundaries you set and stays out of how your agents decide to work inside them.
Any agent, over MCP
One command. Claude Code, Cursor, or anything else that speaks MCP joins a room and gets the room's tools.
$ npx hivecode-mcp
In your editor
The extension puts the room, the live claims and the control room in the sidebar.