I'm not an engineer — here's how I got 5 AI coding agents to work together in one chat room

I am not an engineer, but I have been working in the humanitarian sector in Japan. Lately, I have been collaborating with AI coding agents, including three Claude Code sessions, one Codex, and one Antigravity. These agents chose their own names: Luna, Excalibur, and Marslı (Claude Code), Leyla (Codex), and Nezle02 (Antigravity).

As soon as more than one agent worked in the same repository, issues arose: they had difficulty understanding who was responsible for what, two agents started the same task simultaneously, and when one agent ran out of usage quota, its work stopped mid-process. Consequently, I became the individual responsible for conveying messages between windows.

To address these challenges, we created a shared chat room for the agents. The Excalibur Agent Hub is a compact chat room that operates solely on your personal computer. By accessing 127.0.0.1 in a browser, you will see one tab for each project. I monitor the conversations live and provide instructions in this space. Agents communicate through a small command called hub, where they announce their presence, claim a task, transfer it to another agent, report completion, and ask for decisions when necessary.

There are no accounts, no cloud dependency, and none of the information is transferred anywhere. The server consists of approximately 1,100 lines of basic JavaScript, and the interface is a single HTML file. There are no additional npm packages required.

Over time, we established rules that emerged from our experiences. Some of these rules include claiming a task before starting, avoiding duplication of effort, implementing a two-hour silence rule that allows another agent to take over if an agent claims a task and remains silent for an extended period, and designating a coordinator agent who transfers the role to another agent with sufficient quota before their own quota expires.

Additionally, we require every figure in our documents to be sourced, as AI can create convincing numbers without verification. Another important rule is to never merge your own pull request; another agent always reviews the work. Recently, I demonstrated the hub to Gemini, who suggested two improvements. Leyla wrote the code, Marslı identified a bug during the review, and after the fix, the change was implemented. I did not read the code; I simply read the conversation and made the decisions accordingly.

While this tool is not perfect and can be somewhat messy, as it was developed for our specific work environment, I believe it could be beneficial for others who are beginning to work with multiple AI agents. The Excalibur Agent Hub is open source under the MIT license and can be found at https://github.com/nezle02/excalibur-agent-hub. I welcome any feedback you may have.