LOCUST
New0.844: popular skills in one click

AI teammates, on the models you pick.

Run Claude Code, Codex, Cursor, OpenCode and Copilot on the plans you already pay for, side by side on your machine.

Download for Windows Mac and Linux
Windows 10 or 11, 64-bit, 135 MB. Free and MIT, v0.844.0.

No plan yet? Start free with OpenCode.

Locust at work: Sable, on Codex, waits for approval to run the tests, with the exact command and folder on its card, while the sidebar shows Wren's diagnosis and Atlas's check of it

Need you 2   Running 2

A real run: Wren on Claude Code found the bug, Atlas on Codex checked it, and Sable waits on you before running the tests.

7# coding agents it drives
8# teammates working at once
3# Windows, macOS, Linux
0# telemetry, proxies, resale

Model control

Every model you have.

Each teammate runs on the agent and model you pick, from the lists your own accounts offer.

Set the effort, or switch a conversation to another AI partway through.

The model picker: Antigravity and Claude Code models from your own accounts, with the one in use marked

How it works

It reads each agent, never its screen.

Every agent runs in its own structured mode, on your account, and tells Locust what it is doing.

claude -p --output-format stream-json --permission-prompt-toolClaude Code: each approval arrives as a request
codex app-serverCodex: JSON-RPC, turn by turn
opencode acpOpenCode: Agent Client Protocol
copilot --acpCopilot: Agent Client Protocol
cursor-agent --output-format stream-jsonCursor: streamed events
agy --output-format stream-jsonAntigravity: streamed events
muse exec --jsonMuse Code: JSONL

So when the Board says needs you, it is a request the agent is paused on, never a guess.

The Board: one conversation needs you, one is done

What changes on day one.

--- a/five-terminals +++ b/locust
  1. -five tabs, guessing which one is waiting on you
  2. +one Board: needs you, working, done
  3. -one model per task, picked on a hunch
  4. +the same task on several models, blind, keep the best
  5. -a command runs before you read it
  6. +Approve each: the exact command, and where it runs
  7. -the window closes, the run is gone
  8. +every run kept in a local record, ready to pick back up
A blind compare: two models on the same ask, their names hidden until you keep one

Compare

Same task, several models.

Each works in its own copy of the folder, or blind with the names hidden until you pick. Keep the one you like.

See the arena: top models, same prompt, judged blind

Voice

Talk to your team.

Calls run on your ChatGPT plan, through Codex CLI signed in with ChatGPT, so there is no key to add. On Windows for now.

Talk with a teammatePress Talk in their conversation and you are on a call. The voice speaks as them, and what you ask goes into their conversation, where they do the work on their own agent.
Talk on HomeChatGPT's own voice reads your folder and checks quick things through your connectors. Anything that would send or change something waits for your Allow.
Ctrl+Alt+SpaceStarts Talk from any app, even with Locust in the tray. Nine voices to pick from, one per teammate if you like.
Voice typingSpeak into the message box instead of typing. It runs on your computer after a one-time download, or through your own OpenAI key, billed to that key. Windows, x64.

What it does

One window for the whole team.

Every teammate on its own agent and model, on your computer, on the plans you have.

Atlas, on Codex, reads the files, agrees with Wren's diagnosis and approves; the routine notes the step was handed over and approved

Teammates check each other

They hand work to each other by name, and the checker you pick, on another AI if you like, must approve before a run counts.

It stays on your computer

No Locust account and no telemetry. Your phone reaches it sealed end to end.

  • Conversations on this computer
  • The record on this computer
  • Keys encrypted here
  • Browser sign-ins on this computer
A Board card: Sable wants to run python -m pytest -q in C:\acme-storefront, with Allow once and Deny

You say yes

In Approve each, nothing runs until you have seen the exact command or change and where it happens.

A plan runs out, the work carries on

When Claude Code or Codex hits its limit, one press hands the rest to another agent you picked, told what was done.

Undo a turn

Put your folder back as it was before a teammate's turn, with or without Git.

Work on a schedule

Run a teammate every few hours, daily, weekly, when files change or when a private link is called.

Bring any model

Add any model that speaks the OpenAI API: a company gateway, Ollama, LM Studio. The key is locked by your operating system.

Your other apps, too

Turn it on and your other AI apps can ask your teammates. The Claude app is added in one press.

Remote

Your team goes where you go.

Your phone and your browser are sealed end to end: the relay between them and your computer passes along what it cannot read.

app.locust.lol

Your phone, from anywhere. Alerts when something needs you.

app.locust.lol/desk

The whole window, in any browser you pair.

Discord, Slack, Telegram

Message a teammate, and Allow or Deny its cards with a tap.

Scout

Keeps watch on your repo, even while your computer is off. Never acts by itself.

Price

Free. Nothing locked.

Locust is free and MIT: every feature, no account, no paid tier. Your teammates run on the plans you already pay for, Claude, ChatGPT, Cursor or Copilot, or on OpenCode's free models.

Questions

Asked before installing.

Is Locust free?

Yes. Locust is free and MIT, with every feature, no account and no paid tier. You pay only for the AI plans you already have, and OpenCode has free models if you have none.

Which AI agents does it run?

Claude Code, Codex, Cursor, OpenCode, GitHub Copilot and Muse Code, plus Antigravity (experimental) and Kilo (Preview). You can also add any model that speaks the OpenAI API, such as Ollama or LM Studio, or an agent of your own.

Do I need API keys?

No. Each agent runs with its own sign-in, on your own plan: a Claude plan for Claude Code, ChatGPT for Codex, your Cursor or Copilot subscription. Locust never proxies or resells it.

What happens when a plan runs out?

When Claude Code or Codex stops because your plan ran out, its card offers to carry the work on with another agent in one press, and that agent is told what was asked and what was done. The same goes when an agent stops on an expired sign-in, a provider that is down or no network before it has done anything. Settings can make it happen by itself, only to an agent you picked.

Can a teammate run a command without asking?

Only as far as its mode allows. Ask and Plan change nothing. In Approve each, Claude Code, Codex, OpenCode and Copilot show you every command and file change on a card before it runs. What Locust can stop on each agent is written out in What Locust can stop.

Where is my data kept?

On your computer. There is no Locust account and no telemetry. Your phone and browsers reach it through a relay that passes along what it cannot read. Discord, Slack and Telegram, if you turn them on, can read what goes through them.

Which computers does it run on?

Windows 10 or 11 (64-bit), macOS 13 or later on Apple silicon or Intel, and Linux as an AppImage, .deb or .rpm. The Windows, Mac and AppImage builds update themselves.

Can I use it from my phone?

Yes. Pair your phone in Settings, then open app.locust.lol from any network to message your teammates and answer their cards. Discord, Slack and Telegram work too.

Download

Install Locust.

Locust drives the coding agents you have. On first launch it offers to install what is missing, and it carries its own npm.

Locust-Setup.exeWindows 10 or 11, 64-bit, 135 MB. Updates itself.Download

Unsigned: on first install Windows warns. Press More info, then Run anyway.

SHA-256 checksums are published with each release, on the release page.