Your AI coding agents, as one team on your own computer.
Locust runs Claude Code, Codex, Cursor, OpenCode, GitHub Copilot, Antigravity, Muse Code and Kilo side by side. Each teammate has a name and a role, and runs on the agent and model you pick, under your own accounts. You watch every step, approve what matters, undo a turn you don't like, and keep a record of everything they did. Reach the team from your phone or any browser while Locust runs at home.
Free and open source (MIT). No account needed, no model of Locust's own, no telemetry. Sign in with GitHub if you want a sealed copy of your Locust kept for your next computer.
Download for Windows, Mac or Linux · Releases · What's new · Ready-made teams
If Locust is useful to you, a star on this repository helps other people find it.
The real app (0.845), filmed by Locust's own test drives (_tools/film-launch-clip.mjs).
 Your team of six, each teammate on its own AI and model. Connected accounts: 8 ready. |
 One picker for the models your own accounts offer, grouped by agent, with the one in use marked. |
 A run in progress: a four-step plan, step 1 under way, the live line naming the step and its time. |
 Approve each: before a command runs you see the exact command, where it runs, and what Locust can and cannot tell about it. |
 A hand-off: Wren (OpenCode) diagnosed the bug; Atlas (Codex) read the code himself and wrote VERDICT: APPROVED. |
 The finished fix: 2 files edited, a test file added, four tests passing, in 33 s on a free model. |
 Blind compare: two models on the same ask, names hidden until you keep one. |
 The Board: every conversation by what it needs from you. |
Screenshots from the packaged app on a fresh demo profile (0.838.0; the hand-off and the Blind compare from 0.619.0): real runs, no mocks, taken by _tools/drive-promo-shots.mjs. The models shown are what that machine's accounts offer; yours come from your own accounts.
Skills are folders of instructions an agent follows. Most skill packs are written for one agent. Keep a pack once in Locust and every teammate gets it, whichever agent it runs on.
Open Settings › Teammates › Skills from GitHub. The shelf has the most-starred packs (Matt Pocock's skills, Superpowers, Anthropic's skills, Impeccable, Claude-Mem) and a New this week row of the packs climbing fastest right now (Unslop, tstack, pixelfx, Dimensional Drawing). Or paste any public repository. Locust lists every skill and names every file in it that runs before anything is kept, and keeps the exact commit you looked at.
- Works side by side, on any agent. Up to 64 teammates, 8 conversations running at once, each on its own agent and model. A teammate can hand work to another, and a checker on a different agent can be asked to approve it.
- Asks before it acts. Five modes, from Ask (read-only) to Auto. In Approve each, a card shows the exact command and folder before it runs. Where an agent offers no place to ask, Locust says so instead of pretending (what Locust can stop).
- Keeps a record, and can take a turn back. Every event is written to a local, append-only record before the screen shows it, so a conversation survives a restart. Undo these changes puts back the files a turn changed.
- Keeps working on a schedule. Routines run a teammate every few hours, daily, weekly, when files change or when a private link is called, and carry on in the background after the window is closed.
- Keeps watch. Scout, off until you turn it on, reads your pull requests, checks and reviews and tells you what needs you, with the next step drafted for the right teammate. It never acts by itself.
- Follows you. Pair your phone once, or sign in with GitHub at app.locust.lol/desk, to answer approval cards and talk to your team from anywhere. Messages are sealed on your own devices; the relay in the middle cannot read them. Discord, Slack and Telegram work too.
- Talks. Talk is a voice call with ChatGPT's own voice, through your signed-in Codex. It reads your folder and hands bigger work to a teammate.
- Compares models. The same task on two or three models at once, each in its own copy of the folder. Blind hides the names until you keep one (three models building the same game).
- Shares a team as a picture. Share team › Save image puts your teammates and their routines inside a PNG; Add team from image brings them in, nothing started. Ready-made teams.
- Answers your other apps. A local MCP server, off until you turn it on, lets the Claude app, Claude Code, Codex and OpenClaw ask your teammates. Claude on the web reaches it through the relay, once you allow it on your computer.
| What it is | How Locust differs | |
|---|---|---|
| Claude Code, Codex and the other agents' own apps | One company's agent, on that company's models. | Locust runs them as teammates in one window, next to six other agents, so you don't choose one. Each keeps its own sign-in and its own tools. |
| Orkas | A desktop team led by its own Commander model, which hands parts of a task to built-in specialists and CLI agents. | Every agent you bring leads its own conversation; no model of Locust's own sits above them. No sign-in, and your phone, a browser and chat apps reach the team in the open-source app itself. |
| Orca | A worktree, a terminal and a browser for each task, across many coding CLIs. | Teammates with roles, approval cards before commands, a record of every step and Undo. Built for people who would rather not live in a terminal. |
| Paperclip | A self-hosted web control plane for "an AI company": org chart, budgets, heartbeats. | A desktop app with nothing to host. Your computer runs the agents; the relay only carries sealed messages when you turn it on. |
| CrewAI, LangChain | Code frameworks for building agents into your own program. | Nothing to write. You direct real agents by chatting with them. |
Locust is for you if you already use more than one AI agent, want to pick the model for each job, and want to see and approve what they do.
It isn't for you if you want one chatbot, or a library to build agents into your own program.
You ──► a teammate (name, role, face; its own agent, model and mode)
│
├─ Claude Code ──── the CLI you installed and signed into
├─ Codex ─────────── "
├─ Cursor ────────── "
├─ OpenCode ──────── "
├─ GitHub Copilot ── "
├─ Antigravity ───── the IDE's own agent, through its CLI
├─ Muse Code ─────── "
└─ Kilo ──────────── its free model needs no account (Preview)
│
▼
one append-only record per conversation, written before the screen is
- Agents. Locust finds each installed CLI without reading its credential files, offers only the models that CLI lists, and turns each one's own output into one shape of event. One reader per agent, each built from streams measured off the real CLI: packages/runtime-adapters.
- The record. Every event, card and answer is appended to the conversation's ledger before it is drawn: packages/mission-store.
- Approvals. A card waits for you, or for a rule you saved, and the record names who decided: approval-channel.ts, approval-rule-store.ts.
- Routines. Schedules, file triggers and links: routine-runner.ts.
- Skills from GitHub. Listed before kept, pinned to a commit: skill-library.ts.
- Team pictures. A team inside a PNG: team-card.ts.
- From anywhere. X25519 and AES-GCM between your devices; the relay is a Cloudflare Worker that only forwards: relay-crypto.ts, apps/relay.
Every connection Locust itself makes is listed in docs/NETWORK.md.
Download the latest build at locust.lol, for Windows, Mac and Linux, or from Releases, which has every build's notes and the same files with their SHA-256 sums. The Windows and Mac builds are not signed by a publisher yet (why, and the signing policy), so each asks once:
- Windows warns that the installer comes from an unknown publisher: choose More info, then Run anyway.
- Mac refuses the first open, since Apple has not checked the app. Close that message, open System Settings › Privacy & Security, and press Open Anyway beside Locust.
- Linux comes three ways. The .deb (Ubuntu, Debian, Mint) and the
.rpm (Fedora, openSUSE) install like any package: open the file, or run
sudo apt install ./Locust-*.deborsudo dnf install ./Locust-*.rpm. The AppImage runs on any other Linux and updates itself: allow it to run as a program (chmod +x Locust-*.AppImage), and on Ubuntu 22.04 or later install FUSE first (sudo apt install libfuse2t64, orlibfuse2before 24.04).
Locust updates itself from its own downloads at dl.locust.lol. A .deb or .rpm, which belongs to your package manager, says when a new version is out and hands you the new package.
Requirements: Node.js 22.22+ and pnpm 11.
cd locust
pnpm install
pnpm devpnpm dev starts the Electron app with the React renderer in development
mode. pnpm check builds every package, runs the TypeScript checks and the
whole test suite (about 14,300 tests across the desktop app, the adapters, the
ledger and the relay, as of 0.848).
Green tests are not the evidence here. Each package carries a mutation control
(test/mutation-control.mjs) that breaks one behaviour at a time and requires
the named test to fail. The drives in _tools/drive-*.mjs are the other half:
each launches the built app, does what a person does, and keeps its
screenshots.
apps/desktop: the Electron + React app. The main process (agents, ledger, approvals, routines, updates), the preload bridge and the renderer.apps/relay: the Cloudflare Worker behind phone, browser and app access. It forwards sealed messages.packages/runtime-adapters: finding installed CLIs, the process transport, and one event reader per agent.packages/mission-store: the versioned, append-only ledger and its reader._tools: the release tools and the drives.docs: NETWORK.md, WHAT-LOCUST-CAN-STOP.md, CODE-SIGNING.md, the architecture notes and the assets.
Locust is developed in this repository, and each release is cut from main.
Issues are welcome; code changes are by invitation (see
CONTRIBUTING.md). An agent session working here reads
AGENTS.md, PROJECT.md and docs/CROSS_TASK_CONTEXT.md first.