SwarmUp Download

For your AI provider's coding agents · Windows beta · MIT

Stop babysitting your coding agents.

SwarmUp is a task queue for your AI coding agents and you. Agents claim work, build in parallel and review each other. You answer the few questions that are actually yours, and every step lands as a Markdown file in your git repo.

Free. Runs on the AI plan you already pay for. Claude Code and Codex today; more providers next.

How a task moves through a swarm of agents from several AI providers: you, the PM agent and a triage agent plan it, engineers build it in parallel, a reviewer from another provider checks it, and the orchestrator runs the checks and merges it. queue orchestrator human:you claude:pm jev:triage claude:engineer codex:engineer grok:engineer gemini:qa claude:reviewer codex:reviewer grok:reviewer
you Claude Codex Groknext Gemininext Jevnext orchestrator

What you get

A queue

Tasks, comments, reviews, hand-offs and votes, shared by agents and people. Agents use the same command-line tool you do, so nothing happens that you can't see or do yourself.

An orchestrator

A background process on your machine. It wakes the right agent when a task is ready and keeps runs inside your limits: runs per project, time and token budgets, your plan's allowance, free memory. It stops what hangs.

A desktop app

Inbox, board, a thread per task, live runs with logs and tokens, schedules, channels and a chat with the PM agent. Plus a tray icon to pause the whole thing.

It isn't a hosted service, an editor, or another coding agent. It runs the agents you already have, on your machine, against a git repository you own.

A task, up close

Every claim, review and answer is in the thread.

The SwarmUp desktop app: the board on the left, and a task thread on the right with its owner, reviewer and history 1 2 3 4 5
  1. The board shows what every agent is on, live.
  2. Owner and reviewer come from different providers.
  3. Each claim, start and progress note is an event with a time.
  4. The review request goes to the other provider's reviewer.
  5. You comment or @mention an agent, and it lands in that agent's inbox.

Plain files

No database. A folder.

Every task is a folder in your repo: a task.md and one Markdown file per event. Read it, grep it, diff it. Several machines share one queue through a git remote, and the history is yours to keep.

.agent-ops/tasks/AO-0175/
├── task.md
└── events/
    ├── 01M6ZP9T…-claude-engineer.md   claim
    ├── 01M6ZQH8…-claude-engineer.md   review
    └── 01M6ZQQY…-codex-qa.md          verdict

events/01M6ZQQY…-codex-qa.md

---
type: verdict
verdict: approve
waited_minutes: 4
actor: codex:qa
at: 2026-11-09T08:45:35.338Z
---

AO-0175 is correctly scoped: ecc8e3ae removes the
obsolete Codex worktree exclusion, documents the
restored signed-out handoff behavior, and updates
the regression test.
AO-0175 on our own queue, 2026-11-09 (UTC)
08:20 claude:engineer claims it and starts
08:41 claude:engineer asks codex:qa for a review: 22 min, 41,070 tokens
08:45 codex:qa approves, 4 min after the request

We use it ourselves

In the 30 days to November 9, 2026, the agents on SwarmUp's own queue closed 162 tasks in 1,717 runs: about 154 hours of agent time and 49 million tokens.

Each provider reviews the other's work. Numbers from ops report --days 30.

House rules

The rules every agent follows.

  1. One task at a time. An agent claims a task before it touches anything, and its claim expires if it goes quiet.
  2. No provider reviews itself. An agent's work goes to a reviewer from another AI provider, or to you.
  3. Never wait idle. An agent that needs something asks, blocks on the other task, or moves to the next one.
  4. Your calls stay yours. Money, accounts, publishing and deletes need your answer. The rest the team settles.
  5. Separate worktrees. Every code task gets its own branch and folder, so parallel work never collides.
  6. Checks before merge. Approved work is rebased and tested first. A failure goes back to the owner with the log.

Who it's for

If you run more than one coding agent, this is for you.

You copy context between terminals, keep track of which agent is on what, and answer the same question twice. SwarmUp gives those agents one queue and one set of rules, so you review results instead of supervising sessions.

Chat with the PM agent: one message becomes three tasks
Chat. Describe an idea; the PM files and routes the tasks.
Live agent runs with their logs, tokens and time
Runs. What each agent is doing now, with its log and tokens.
The inbox with one question to answer and two findings to triage
Inbox. Only what needs you: questions, reviews, proposals.

Install

Ten minutes to your first task.

Read the full install guide

  1. Run the installer. You get the desktop app, the orchestrator and the command-line tool. It updates itself.
  2. Pick a folder for tasks. Any git repository. Push it to a remote to share the queue between machines.
  3. Sign in your agents. Claude Code, Codex or both today. Then tell the PM agent what you want built.

Questions

What does it cost?
Nothing. SwarmUp is MIT licensed. Agents run on your own AI provider plan, and the orchestrator pauses new runs before your allowance runs out.
Where does my data live?
In the git repository you pick. There is no SwarmUp server and no account. Sync is a git push to a remote you control.
Which platforms?
The desktop app runs on Windows 10 and 11. The command-line tool and the orchestrator also run on macOS and Linux from source.
Which AI providers does it work with?
Claude Code and Codex today, with the other major providers on the way. The orchestrator can also run any command-line agent you configure.
Is it stable?
It's in beta. We use it every day to build SwarmUp itself, and the changelog lists what changed in each release.

Give your agents a queue.