Coordinate Claude Code, OpenAI Codex and Pi into collaborative engineering teams. Zero background daemons, zero database locks, 100% native CLI.
npx @team-agent/installer@latest installOpen source (AGPL-3.0) · macOS & Linux · uses the agent CLIs and subscriptions you already have
$ team-agent claude Claude Code started as the leader of this workspace > Build a team for this repo: Codex implements the billing API, Pi writes the tests, and a second Claude reviews every change before merge. ✓ wrote .team/current/TEAM.md and 3 role files ✓ started 3 teammates in their own tmux panes $ team-agent status name provider activity health backend codex working ok tests pi working ok reviewer claude idle ok $
Team Agent message from leader: Add pagination to GET /v1/invoices and cover it with tests. • Reading src/billing/invoices.ts • Editing src/billing/invoices.ts +48 −6 • Editing tests/invoices.test.ts +61 • Running npm test -- invoices ✓ 14 passed (1.8s) → report_result to leader status: done · 2 files changed · tests passed ›
Team Agent message from backend: Pagination is in. Please run the integration suite. • Writing tests/invoices.e2e.test.ts • Running npm run test:e2e ✓ 9 passed ✗ 1 failed limit=500 should be rejected with 400 → send_message to backend "GET /v1/invoices?limit=500 returns 200, expected 400" ›
team-agent doctor output when everything is healthyRunning more than one coding agent on a project sounds like free parallelism. In practice it breaks down in the same few ways for everyone.
Claude designs, Codex implements, and you shuttle context between windows all day. The agents never talk to each other, so you do it for them.
A closed terminal, an SSH drop or a reboot kills the agents and their context. Tomorrow you re-explain the whole project from scratch.
Many orchestration tools add an always-on server and a database queue. When it stalls or runs an old version, every agent stops, and memory stays allocated even when nobody is working.
You keep working in Claude Code, Codex or Pi exactly as before. Team Agent adds the missing layer: starting teammates, routing messages, and keeping everything alive and recoverable.
Pick the best model for each role: Claude for planning and review, Codex for fast implementation, Pi for any model catalog you have access to. Every teammate keeps its own provider, login and settings. No vendor lock-in, and you stack the subscriptions you already pay for.
Every agent runs in tmux, so closing the terminal or losing SSH changes nothing. After a reboot, team-agent start brings the team back with each agent's conversation resumed. Need a clean slate for one teammate? Reset it while the rest keep working.
There is no always-on server and no database. Each command does its job and exits. Messages live in small append-only inbox files you can read with tail. Stop the team and Team Agent leaves nothing running: no daemon, no memory, no open database.
Team Agent types into each agent's own terminal (PTY) through tmux, the way you would. It runs the official binaries from your PATH with your configuration and wrappers. Your tools are not intercepted; agents simply gain three small tools to message teammates and report results.
Everything runs on your machine, in processes you can see. There is no hidden service between you and your agents.
Say “build a team” to your leader, or write a short TEAM.md plus one file per role with its provider and instructions.
Each teammate starts in its own pane. Messages are saved to the recipient's inbox first, then typed into its terminal, so nothing is lost if an agent is busy.
Teammates report results to the leader, which summarizes and only asks you when a real decision is needed. Stop for the day; resume tomorrow.
Mix providers freely. The leader and every teammate choose their own CLI, model and permission settings.
Third-party and self-hosted models (for example DeepSeek through an Anthropic-compatible endpoint) can be attached with a local profile. Keys stay in local files and never enter prompts.
Short walkthroughs of real team runs.
Requirements: macOS or Linux (WSL works), Node.js 18+, tmux, and at least one agent CLI installed and signed in.
One command installs the team-agent binary for your platform.
npx @team-agent/installer@latest installUse team-agent claude instead of claude. Works the same with codex or pi.
team-agent claudeDescribe the roles in plain language. The leader writes the team files, starts everyone and reports back.
Build a team: Codex implements, Claude reviews.The honest alternatives are a single agent with built-in subagents, or several terminals you coordinate yourself.
| Capability | Built-in subagents | Terminals by hand | Team Agent |
|---|---|---|---|
| Mix Claude Code, Codex and Pi | No, one provider | Yes, manually | Yes, per role |
| Agents message each other | Report to parent only | No, you relay everything | Directly, plus results to the leader |
| Teammates persist across days | Usually one-shot | Until you close the window | Yes, resume with context |
| Survives terminal close or reboot | No | No | Yes (tmux + session resume) |
| Background service required | None | None | None, 0 MB idle |
| Uses your existing subscriptions | Yes | Yes | Yes, several at once |
Practical guides for developers running AI coding agents locally.
Why pairing a planner-reviewer with a fast implementer works, and a step-by-step setup in tmux.
Always-on servers and database queues fail in predictable ways on developer machines. What to use instead.
How to send messages to interactive CLIs like Claude Code and Codex without losing or duplicating them.
Team Agent is an open-source command-line tool that turns terminal AI coding agents such as Claude Code, OpenAI Codex and Pi into one collaborating team. You talk to a single leader agent; it starts teammates in their own tmux panes, sends them tasks, and collects their results. Teammates can message each other directly and keep their memory between sessions.
Install Team Agent with npx @team-agent/installer@latest install, start Claude Code as the leader with team-agent claude, and ask it to build a team, for example: a Codex teammate that implements and a Claude teammate that reviews. Team Agent launches each CLI in its own tmux pane with your existing logins and routes messages between them. Read the full guide →
No. Team Agent is a plain CLI with no background server. Each command does its work and exits. The only Team Agent processes you will see are attached to agents you are running (your leader's launcher and a small messaging helper per agent), and each exits with its agent. Close your agents and, with nothing scheduled, nothing remains: a 0 MB idle footprint. Scheduled reminders use a short-lived process supervised by the operating system (launchd on macOS) that exits when the schedule is empty. Why we built it this way →
Closing the terminal does not stop the team, because every agent runs inside tmux. After a machine restart, team-agent start brings the team back and resumes each agent's saved conversation where the agent CLI supports it, so teammates continue with their previous context.
No. Team Agent runs the official agent CLIs you already use, signed in with your existing subscriptions or accounts. If you prefer API access or a third-party model, you can attach a local profile with its own endpoint and key; keys stay in local files and never enter prompts.
Built-in subagents run inside one session, report only to their parent and are usually gone when the task ends. Team Agent teammates are persistent processes with their own conversation history; they can use different providers, message each other directly, and are still there tomorrow.
macOS (Apple Silicon and Intel) and Linux x64, including Windows through WSL. You need Node.js 18 or later, tmux, and at least one supported agent CLI.
Yes. Team Agent is open source under the AGPL-3.0-or-later license. You only pay for the agent subscriptions or API usage you already have. Commercial licensing is available on request.
Install once, keep using the agents you know, and let them work together.
npx @team-agent/installer@latest install