T3 Code already runs every coding agent you own: Claude, Codex, Cursor, side by side in one inbox. The glaring hole: no native agent of its own, trained to orchestrate that fleet on your behalf. That's what this is. Native, built in, running out of the box on a single API key, on your model, or entirely on your own machine. And it runs the rest of them for you.
Watch it run
One prompt. The agent fans six questions out across Claude and Codex, reads the answers back, follows up, files the threads, and announces every move out loud through the Mac's speech engine while it works. Nobody touches the keyboard after Enter.
100 seconds, one take·
six live threads across two providers·
the voice is say, run by the agent itself·
the whole thing ran on DeepSeek V4 Flash for under a penny
The headline act
T3 Agent isn't another entry in the provider list. It operates the list. It knows which agents you have, what each costs, and what each is good at. You say what you want done; it decides who does it, launches them in parallel, and reads their work back.
the app's own dark theme, tokens and all·on loop·the real thing is the video above
It can see which providers are configured, whether each one bills per token or draws on a subscription you've already bought, and exactly which models each will accept. So it can put the cheap work on the plan you're already paying for.
Provider, model, and how hard to think, chosen per task. Ask for a reasoning level a model doesn't offer and it uses the nearest one that model does, instead of refusing and wasting the round trip.
Not the summary a thread wrote about itself. The real transcript, with the failed tool calls marked and the errors counted. If a delegate is stuck waiting on a question, it can see the question and answer it.
Send a correction back into the same thread, rename it, snooze it, pin it, settle it or archive it. The inbox stays yours, but you don't have to do the filing.
Every delegated thread gets its own git worktree and branch, so five agents can work the same repo at once without overwriting each other. You merge what you like, when you like.
Tell it to wait on a delegate and it holds on a real server-side wait (no polling loop) and wakes the moment the thread settles or stops to ask a question.
Before
You are the message bus.
After
You supervise a swarm from one chat.
Also in the box
A dispatcher is only as good as its own hands. Before T3 Agent orchestrates anything, it does the ordinary job properly. Not a chat box bolted onto the sidebar, but a first-class in-process provider, so every feature T3 Code already has works on it unchanged. Paste one API key and it's coding straight away, on models priced like DeepSeek V4 Flash, the one the demo runs on.
Reads, writes, edits, globs, greps and runs commands, with path safety that keeps every one of those inside the project.
Obeys the thread's runtime mode and the same approval prompts as every other provider. The gate is a pure function, tested on its own.
Picks up AGENTS.md and CLAUDE.md, loads skills, and connects the
MCP servers you've configured.
Every turn is a hidden git ref, so you can rewind a thread exactly the way you already do with Codex or Claude.
Fans work out to parallel sub-agents when a task branches, and compacts the conversation when it gets long.
Every thread carries a live cost readout: total spend and what the last turn cost, priced from your provider's own published rates, cache discounts included. A model it can't price reads "unknown" instead of pretending to be free.
Try it, right here
The picker, exactly as it ships, in dark. The whole frontier filed under one provider: yours.
Bring one key
These ship and work; they just have fewer miles on them than OpenRouter does.
Under the hood
No hidden capabilities and no marketing verbs. This is the actual tool surface.
Working on your code
Running your other agents
Read this part
Stated up front, because finding out later is worse.
It can hold on a delegated thread until the work settles: a real server-side wait, not a polling loop. But nothing pings it unprompted: there's no background watcher, and it will tell you that rather than promise to watch things it can't see.
Not threads, not projects. Deletion is the one action with no undo, so it isn't wired up.
Which means it can never widen its own reach by raising a delegate's permissions.
Not unless you turn that on separately. Approvals are how you stay in the loop.
Nothing per turn on a subscription; real money on an API key. There's no cap on how many run at once, so a wide fan-out spends widely.
Upstream says to expect bugs. This adds twenty-two thousand lines of code on top of that. Treat it accordingly.
Get it
Installs as T3 Code (T3 Agent), with its own name and its own bundle ID, so it lands beside your real T3 Code, never on top of it.
arm64 for Apple Silicon, x64 for
Intel ( → About This Mac if unsure). Open it, drag to Applications. Your existing
T3 Code is untouched.settings.json.One caveat worth knowing: it shares
its on-disk data (~/.t3) with an official install: same projects, same threads,
same provider keys. Run one at a time.
No prebuilt binary yet; build it from source. Needs Node 22.16+, 23.11+ or 24.10+.
# clone the branch git clone -b feat/t3agent-upstream-pr \ https://github.com/BenSyne/t3code.git cd t3code # install vp, skip if you have it curl -fsSL https://vite.plus | bash vp i vp dev
npx t3@latest installs upstream, which does not include the
agent. The only ways to get this build are the download above or the source clone beside it.
Everything upstream does, this still does. It's upstream with one provider added, not a rewrite.
Dear T3 team
First, the important part: this is a tribute. T3 Code is my favorite tool in the category, the best multi-agent command center anyone has shipped, and I use it every single day. I built this because I love the product, not because I think you're doing it wrong.
You've said thousand-line PRs are a good way to get banned. This one is twenty-two thousand, so it lives on my fork: full diff, comments open, nobody's inbox harmed. Your README says to fork the editor you want. So I did. Anything useful is MIT: take it.
Built on upstream 89c320df · v0.0.33 · 11 Aug 2026