say lo.your agents get to work.

One chat for every coding agent on your machine. Text it, talk to it or tap from your phone: it starts the sessions, routes the work and only asks you when it matters.

$ curl -fsSL https://raw.githubusercontent.com/agtt/orchestra/main/deploy/install.sh | bash

A fresh Ubuntu 24.04 server, as root. About 10 minutes.

lo2 working
needs you · api

Run the migration on the live database?Claude Code · step 4 of 6

Sessions
  • Fix the checkout bugshop · Claude Code · working
    4m
  • Add search to the blogblog · Codex · working
    1m
  • Write the API docsdocs · OpenCode · done
    ✓

Try it: tap Allow, or tell it something.

Drives the agents you already pay for

Claude CodeCodexOpenCodeAntigravity Gemini ACPCursor ACP
Cline ACPKimi ACPGoose ACPQwen ACP Ollama localLM Studio local

An account hits its limit? lo hands the task to the next one, then to a free model, then to one running on your machine.

How it works

Three steps. Then just talk.

01

Install

One line on a Linux server, or bin/setup on a Mac. It asks for your Telegram bot token and a Claude sign-in.

02

Pair your phone

Open the address it prints, tap Pair, tap Allow on Telegram. After that, Face ID or a passkey signs you in.

03

Say “lo, …”

“lo, fix the checkout bug.” It starts a session, picks the agent, and pings you when it’s done or needs you. Alerts cost zero tokens.

From anywhere

Your phone is the remote.

The app

Sessions, approvals, files and routines on one screen. Add it to your home screen; Face ID signs you in.

Voice

Talk instead of type. Speech is transcribed on your own machine, so the audio never leaves it.

Siri

“Hey Siri” a task, ask for status or approvals, and hear the answer back. Shortcuts are made for you in the app.

Telegram

Your own bot. Text or voice notes in, results and Allow / Deny buttons out.

WhatsApp

Link your numbers. lo reads chats into work lists and writes only what you tell it to.

Beyond code

It has hands, too.

Your agents can work your browser, your Mac apps and your social accounts, and hand you anything that needs a human.

Browser use

The lo Chrome extension drives your own Chrome, already signed in to your sites.

  • Each session gets its own tab group, in a separate lo window
  • Never takes your window; you keep browsing
  • CAPTCHAs, passwords and codes are handed to you

Computer use

Any Mac app, not just the browser. It reads apps through the macOS accessibility tree.

  • Presses buttons, fills fields and picks menus in the background
  • Leaves your pointer and keyboard alone, so you keep working
  • Each step on your phone to allow, or a 10-minute pass
  • Several agents at once, lighter on tokens Soon

Social media

Connect your accounts and post straight through each platform’s API. No browser, no extra server.

  • Instagram, YouTube, TikTok, X, LinkedIn and 15+ more
  • Any number of accounts; see what’s connected
  • Post now or at a set time

Runs where you do

Your machine. Your keys.

MacHomebrew + bin/setup
LinuxOne line in a terminal
CloudAny VPS, ~10 min, your own domain with passkeys
WindowsThrough WSL2, as another machine

Many machines, one chat. Add a laptop, a server or a Windows PC with one line; lo picks where each task runs.

Safety

It just does it. Except for these.

No Allow / Deny chasing. lo starts, pushes and deploys on its own, and stops to ask only before:

  • Destroying data
  • Deleting folders
  • Force-pushing over history
  • Spending money
  • Touching secrets and keys

A guard enforces it, not a prompt. Web pages and other sessions’ output are treated as data, never as orders.

FAQ

Fair questions.

Is lo free?

Yes. lo itself costs nothing. You bring the AI plans you already have: a Claude account, ChatGPT for Codex, a Google account for Antigravity. Free OpenCode models and local models work too.

Is it open source?

The code is going public on GitHub soon. Until then the GitHub link is where it will live.

Which plans and agents does it use?

Claude Code, Codex, OpenCode and Antigravity, plus ACP agents such as Gemini, Cursor, Cline, Kimi, Goose and Qwen, and local models through Ollama or LM Studio. You can add several accounts per tool; when one hits its limit, the task moves on by itself.

Does my data stay local?

lo runs on your own Mac or server, with your own logins and your own Telegram bot. Nothing is shared between installs and there is no lo cloud in the middle. Voice is transcribed on your machine. Your prompts go only to the model providers you choose.

Can it use my browser and my Mac?

Yes. Browser work runs in your own signed-in Chrome through the lo extension, in a window of its own, so yours stays yours. Computer use works any Mac app in the background through the accessibility tree, and each step shows up on your phone to allow. Passwords, codes and CAPTCHAs are always left to you.

Can it post to my social accounts?

Yes. Connect Instagram, YouTube, TikTok, X, LinkedIn and more, then ask lo to post now or schedule a post. It posts through each platform’s own API, and only what you asked for.

Do I need a Mac?

No. It runs on a Mac, on a Linux box or on a cloud server, and a Windows PC can join through WSL2.

say lo.

Your agents are waiting.