Bro vs T3 Code.

The short version.

Pick T3 Code for one calm window that treats every agent the same and lives in git. Pick Bro for the agents themselves, on any model, with logins handled and a crew that keeps working while you're away.

What each one is.

Both are free, open source and built around the same idea: you already pay for good coding agents, so the interesting problem is running several of them well.

Row by row.

The same questions for both tools. A filled dot is a yes, a half dot is partly, a dashed one is a no. Flip the switch to hide every row where you'd get the same answer from either.

Managed login. Sign in once.

Every serious agent setup ends up with more than one login: a work Claude, a personal one, a ChatGPT plan, maybe Gemini. Most tools treat that as your problem. You sign in from a terminal, you point the app at the right config folder, and you find out a login ran dry when the agent stops mid-task.

Bro manages the logins for you. You add one from the launcher, Bro runs the sign-in and keeps it in a profile of its own, and from then on any session can run on it. The sidebar shows the 5-hour and weekly limits left on each login. In proxy mode Bro sends new conversations to the login with the most room, and when one hits its limit the same request moves to the next, before the agent has seen an error.

How T3 Code handles it

T3 Code has a good start on this. Codex connects with ChatGPT inside the app, extra ChatGPT accounts sit in the model picker, and Cursor and Antigravity sign in from the app too. Claude and the rest still need their own CLI logged in on the machine, and a second Claude account means a second CLAUDE_CONFIG_DIR and a claude auth login in a terminal. Its FAQ is direct about why it stops there:

For a bigger pool it points to CLIProxyAPI, a separate self-hosted proxy, and shows the pooled limits once you connect one. Since 0.1.5 new threads also skip a provider that just hit its limit, so it's moving this way. Bro has the whole piece built in, so there's nothing extra to run.

Many projects at once.

Real work is rarely one repo. A normal day is a mobile app, an API, a marketing site and a scripts folder, each with an agent or two running, and the hard part is knowing which of them needs you.

Bro groups every session by project, the git root or the folder when there's no repo, and rolls the state up to the project's row, so you see live counts and who's waiting without opening anything. Alt A jumps to whoever needs you, Alt Shift T tiles everything at once, and the phone app groups its chat list the same way.

Coordination follows the same shape. Each project can have its own orchestrator, and a root orchestrator above them opens or reuses those and takes your voice, your phone and your schedule: say "open a session in justgains" and it finds the project. Bro also keeps projects apart. Claude Code lets any session on a machine find and message any other, so a session in one repo can be interrupted by work in an unrelated one, and Bro scopes that to the project by default.

T3 Code's default sidebar is one list of threads from every project, ordered by when each last came back to you, with a collapsed Working section for the busy ones. Grouping by project is a beta setting in Labs, and in that mode threads can't be dragged and the Working section turns off. Search is good: Ctrl K finds threads across connected machines by title, branch, pull request or project. What its docs don't describe is a coordinator for each project or agents scoped to their own. Its MCP server lets one outside agent start threads in every project instead.

A terminal UI and a desktop app, for every use case.

Bro is one workspace that opens two ways. The desktop app draws the terminals on the GPU. bro-cli draws the same workspace in your terminal, with the same sessions, sidebar, launcher, views and keys. Over SSH, in tmux or on a machine with no display, plain bro opens the terminal version by itself, and bro --view opens more windows on whatever is already running.

So the answer to "where do I run my agents" is the same everywhere: at your desk, on a server, on a laptop shut in a bag, from your phone. Only a few things need a window, such as focus mode and the media panel. The sessions, the sidebar and the keys are the same in both.

Any model in any agent.

Bro's proxy translates between the Anthropic Messages, OpenAI Chat, OpenAI Responses and Gemini APIs as requests fly by. Claude Code can run on GPT, Gemini, Kimi, DeepSeek or Grok, and Codex, Pi and omp can run on Claude, with tools, images and thinking coming through intact. There are 16 providers in the catalogue, OpenRouter alone adds 463 live models, and Ollama and LM Studio cover the ones on your own machine.

T3 Code covers part of this by hand. You can add a provider instance with your own keys and a custom endpoint, and its Claude guide has an OpenRouter recipe. What it won't do is local models. Its FAQ answers the question in one line:

If you disagree, or you want a private model on a private repo, T3 Code isn't the tool.

The real agent, not a rebuild.

T3 Code's FAQ describes its goal in one line: "Every harness feels nearly identical in here." That's a design choice, and a good one if you want a single interface. It also means T3 Code draws its own view of each agent from the SDK or protocol underneath, so what you see is T3's rendering of what that layer exposes.

Bro goes the other way. Each session is a real terminal, with kitty keyboard support, ligatures and 10,000 lines of history, so Claude Code and Codex look and behave exactly as they do on their own, and a new release works the day it ships. When you'd rather read than watch a terminal, Alt F flips every Claude Code and Codex session into focus mode: big clean type, one quiet line for each run of steps, questions and permission prompts as buttons, and a model and thinking picker in the footer.

One practical side effect, straight from T3's FAQ: T3 Code drives Claude through Anthropic's Agent SDK, and the FAQ says Agent SDK usage drains your 5-hour limit faster than Claude Code does, a problem it says it has reported to Anthropic. Its tracking issue measures roughly 1.6 to 1.7 times. Bro launches the claude binary itself, so that path never comes up.

The crew behind the window.

Bro's orchestrator is a session that runs the others. It opens sessions, hands out tasks, waits for the answers and tells you when it's done. You can ask it out loud (hold Right Win, say "open a session in justgains"), from your phone, or on a schedule: in 10m, at 17:30, daily 09:00.

Under it are 26 MCP tools that any shell can call as bro ctl, and watchdogs. Press Ctrl Space W on a session and list its milestones. When the agent goes idle, a fresh reviewer reads the chat and the code, checks the work against the milestones and sends the next task.

T3 Code has real answers here too. A thread can delegate to child threads on any provider, an MCP server lets outside agents (Claude Code or Codex in your own terminal, even ChatGPT) start and message threads, scheduled tasks fire on an interval, at a fixed time or from a webhook, and /goal keeps a Codex or Claude run going until its goal is met. What Bro adds is a second pair of eyes: a watchdog reads the chat and the code, checks the milestones you set and hands the agent its next task, and the whole crew answers to your voice.

Your phone, your server, your sessions.

This is T3 Code's strongest ground, and we'll say so plainly. Its iOS and Android apps are on the App Store and Google Play. They send push notifications when an agent finishes or needs approval, show Live Activities on iPhone, and connect through T3 Connect, SSH, Tailscale or a pairing link.

Bro's phone app pairs with Bro on your computer over your own network or Tailscale. It lists every session like a chat, with a spinner while it works, an alert when it needs you and a check when it's done. The terminal is a tap away with its key row, dictation runs on the phone, and tapping a file path opens the file at the right line. It isn't on the stores yet, so for now T3 Code is the one you can install tonight.

Both are built so work outlives the window. Bro's sessions live in a session server, so closing Bro, updating it or crashing it doesn't end them, and the next Bro picks them up, on Windows too. T3 Code can run as a background service on Linux and macOS. On Windows its docs say background services aren't supported, so the desktop app or t3 serve has to stay running.

Where T3 Code is ahead.

Plenty. If any of these is your daily work, take them seriously.

• Git, end to end. Branch and worktree controls in every thread, pull requests created and merged in the app, review guides, files marked as viewed, linked pull requests and GitHub stacks, across GitHub, GitLab, Bitbucket, Azure DevOps and Forgejo. Bro leaves git to your agents and your tools.
• Phones and remote machines. Store apps rated 4.9 on iOS, push alerts, a web client at app.t3.codes and T3 Connect for reaching a box from anywhere.
• Breadth. Nine providers are first-class, including Cursor, Grok Build, OpenCode and Antigravity. Bro has four, plus a plain shell for everything else.
• Fan-out and delegation. Pick several models in the composer and each gets its own thread and worktree, and a parent thread can hand work to a child on any provider.
• More in the window. Shared browser tabs your agents can drive, an iOS Simulator and Android Emulator panel, viewers for PDFs, Office files and notebooks, and agents that watch a pull request and wake on a failed check.
• Momentum. Over 32,000 stars, 60 stable releases and 641 nightlies in about eight months. Bro is young: the Rust rewrite started on September 29 and the desktop app landed on October 21. T3 Code, for its part, still labels itself Beta.

Which one, then?

Match the line that sounds like your week.

You don't have to choose. Both sit on top of the agent CLIs you already have installed, so they don't fight. Run T3 Code for review and Bro for the crew, and see which one you reach for.

Questions people ask.