GitHub Next, the GitHub group that builds experimental developer tools, has released Ace, an open-source project that lets several people and their AI coding agents work in the same conversation while everything runs on their own computers. The project page at ace.githubnext.com calls it an experiment in letting many people steer coding agents at once (agents being programs that take actions on their own), licensed under MIT, and warns readers to expect rough edges.
The pitch is easiest to follow through the demo on the page. A developer named Nate asks the group how to give a music app a dark mode. Two teammates answer in the chat, and Nate types “@ace do it”. The agent reads the code, edits the files, runs the app, and asks the group whether the pads should glow or dim. Any teammate can answer, and the agent picks the work back up.
Ace organizes this around a few pieces. A channel groups related chats and runs as a single worker on one machine. Each piece of work gets a lane, which is its own Git worktree and branch, and only one chat can write to a lane at a time. Finished work arrives as a pull request. GitHub Next says Ace itself is built this way.
Teams are defined by Tailscale, the private networking service. Everyone on the same Tailscale network who runs Ace is on the team, and channel traffic travels directly between their machines over that network. A small shared directory, which the team deploys to its own Cloudflare or Fly.io account, only records which projects and channels exist and which host has them. According to the page, it holds no messages, code, or diffs. A cloud channel, still in beta and optional, can live in the team’s own account and keep running while every laptop is closed.
Model calls go from the user’s Mac to Anthropic or OpenAI with the user’s own key, which stays in the keychain. GitHub Next says no Ace service sits in the middle, and that Ace operates no servers for the team at all. For a company nervous about routing source code through another startup’s backend, that is a real selling point.
The same design carries the risk, and GitHub Next states it plainly on the page. Whatever permissions the person running the host has, the agents have too. They can run shell commands and edit files well outside the project, and the page says plainly that neither channels nor lanes work as a sandbox. A lane keeps a branch tidy; it does not stop an agent from touching anything else the user can touch.
Teammates can only prompt an agent after its owner switches on agent access for that channel, and it starts off. Switching it on means a colleague can run commands on the host. GitHub Next’s advice is to work this way only with people you would trust with the machine itself, to bind the host’s gateway to the local loopback address or the private network only, and to keep it off the public internet entirely.
Ace is a project page and an experiment, so it comes with no benchmarks, adoption figures, or independent security review, and GitHub Next has not said how it plans to support it. What it does offer is a clear model of the trade: a shared agent workspace that is cheap to run and private by default, in exchange for treating every teammate with agent access as someone who can sit at your keyboard. Anyone trying it this quarter should pick a spare machine for the host and decide who gets that switch before the first demo.
Reported by GitHub Next, from the Ace project page at ace.githubnext.com (no publication date given).
