Multi-agent Claude Code you can watch and steer.
A team of Claude Code agents working side by side in tmux. A manager hands work to specialists, they talk over a message broker, and each one can spawn its own subagents. You attach to the session to watch, and drop a message in to steer. Built in two days; no black boxes.
In one paragraph.
Claude Colony spins up a team of Claude Code agents that work a coding task together. A manager agent plans and delegates; specialist agents (UX, frontend, backend, devops) take the pieces and run in parallel. Each one is a full Claude Code instance in its own tmux pane, with the full toolset, and each can spawn its own subagents when a job needs splitting further.
The agents coordinate by passing messages through a file-based broker, addressing each other by name: @frontend, @backend, @all. You attach to the tmux session to watch every pane at once, and steer the whole colony by typing into a shared pane. Ready-made team presets take you from nothing to a working colony in one command.
What makes it different.
Every agent is a full instance
A colony member is not a sandboxed sub-task with a narrow brief. It is a complete Claude Code instance with the full toolset, free to read, write, run commands and spawn its own subagents. The result behaves less like a script and more like a real team, where each member can think for itself.
You can watch and steer
Because every agent runs in its own tmux pane, you see the work as it happens, side by side. Attach to watch, detach without stopping anything, and redirect the whole team by dropping a message into the shared pane. The orchestration runs in the open, not hidden behind an API.
The features that matter.
@mention message broker
A file-based channel where agents address each other by name (@frontend, @manager, @all) to hand off work, ask for help or flag a blocker.
tmux as the cockpit
One pane per agent. Attach to watch the team work in real time, switch between panes, and detach without stopping a thing.
Ready-made teams
Presets spin up a working colony in one command: fullstack, frontend, api, mobile and docs teams, each with sensible roles.
Concurrency safety
File locks and an ownership registry stop parallel agents editing the same file at once and clobbering each other's work.
How it's built, and the system around it.
Colony is a TypeScript CLI that leans entirely on tools that already exist. It drives tmux for the panes and sessions, builds each agent on Claude Code's native agent definitions in .claude/agents, and keeps all coordination state as plain files under a .colony directory: messages, shared channels, locks and review workflows. No server, no database.
That was the whole bet: that genuine multi-agent behaviour could come from wiring together the CLI, the file system and Claude Code, rather than from a heavyweight framework. Two days of building got it working.
- TypeScript CLI · Node 18+
- tmux Panes · sessions
- Claude Code Native agents
- File-based IPC @mention broker · state
The motivation.
In December 2025 I went to the first Claude Code Budapest meetup. An engineer from Anthropic described an internal system they call the Swarm: many Claude Code instances wired together so they can talk to each other, coordinate, and run on their own, with every agent able to spawn more agents. The idea that you could superpower the whole system by letting agents collaborate, rather than scaling one agent, stuck with me.
So I spent the next two days building my own version to see how far I could take it. Colony was the result: that swarm idea made into something I could actually watch and control from the terminal.
Two days to build; an industry too fast to keep up with.
The build itself was the easy part. A working multi-agent orchestration system in two days, mostly by describing what I wanted, is a fair measure of how good AI has become at hard technical problems. When building is this cheap, the interesting question stops being can I make it and becomes is it worth maintaining.
For Colony the answer turned out to be no. Within weeks, Claude Code, Anthropic and others shipped features that covered much of what it did, and I could get the same results from the native tooling. Keeping an open-source AI project current, let alone ahead of a field moving this fast, is a relentless and near full-time effort, so I switched to the newer tools and parked Colony rather than carry it.
It also left me with more respect for the command line than I went in with. Underneath the agents, the whole thing is tmux, the file system and a CLI, and that turned out to be plenty. Sometimes the terminal is the platform.