Agency

Many agents.
Many branches.
One window.

Agency is a free, open-source desktop app that runs AI coding agents in parallel on your own machine. Claude Code, Codex, Gemini CLI and the other agent CLIs you already use each get a real terminal, their own git worktree and their own branch. You read the diffs and merge the good ones.

Download for macOS and Linux Source on GitHub
  • v0.2.5, in beta
  • macOS 11 or later, Apple Silicon
  • Linux, x86_64 and arm64
  • Free and Apache-2.0
  • No account

Bring the agents you already pay for. Agency finds what is on your PATH and offers to install what is not. Your keys and subscriptions stay as they are.

  • Claude Code
  • Codex
  • Gemini CLI
  • OpenCode
  • Cursor
  • Copilot CLI
  • Kimi Code
  • Crush
  • Hermes
  • Pi
  • DeepSeek Harness
  • any other CLI

From an issue to a merged branch

Write the task down, give the agent the context, let several run at once, and read every change before it lands on your base branch.

1 plan

Your tracker is a folder of markdown files

Issues live in your project as plain files. That is why an agent can work them: picking one up, changing its status, commenting and filing a follow-up are all just edits.

  • Dispatch an issue to an agent and it moves to In progress
  • Open a PR and it moves to In review. Merge and it closes
  • Comments, links, attachments and due dates, all in the file
The Issues tab: a board grouped by Backlog, Todo, In progress and In review, with an issue open beside it showing its acceptance criteria, a comment and a linked issue.
2 context

Notes your agents can actually read

A wiki that lives in your repo, with wikilinks, backlinks, tags, an outline and a quick switcher, plus a journal with daily notes and a weekly note built from what actually happened.

  • Hand a note to an agent and watch the page update as it writes
  • Put an agent in the side panel of the note you are working on
  • Keep a notes-only workspace with no code in it at all
The Docs tab: an architecture note open with its properties, headings and a task list, the notes tree on the left, and the note's outline and backlinks on the right.
3 run

Six agents, six branches, no collisions

Every run opens in its own git worktree: a second checkout of the same repo on its own branch. Agents working the same project never touch each other's files. Watch them all in a grid, or pull one full size when it needs you.

  • Different agent CLIs on the same repo at the same time
  • Plain terminals alongside, in the project checkout, when you want one
  • Desktop alerts when an agent finishes, stalls on a question or hits a conflict
The Agents grid: six agents working one repo at once, Claude Code, Codex, Gemini CLI, OpenCode, Cursor and Pi, each tile showing its branch, its diff size and its live terminal.
4 review

Read the diff before you trust it

Per-hunk diffs, commit, history, branches and stash for whichever worktree you are standing in. Review a pull request, open one, merge it, and resolve the conflicts when two agents reach for the same lines.

  • Per-run diffs, so you review one agent's work at a time
  • Comment on a diff and send it straight back to the agent
Source Control Changes: a side-by-side diff of an agent's fix to a flaky test, the commit form for its branch, and the branch history below.

No account. No vendor backend. No telemetry.

Agency does no first-party data collection, and there is no server of ours for it to talk to. The agents are third party and talk to their own providers, the same as when you run them in a terminal.

The one request Agency makes on its own is a check against GitHub for a newer release. It sends no identifiers, and one switch in Settings turns it off.

Read the privacy page

The parts you notice on day two

Terminal first

Agency does not reimplement your agent. It spawns the CLI you already use in a real terminal, TUI and all, and you can type into it.

Agent agnostic

Eleven agent CLIs built in, and any other you can launch from a command line. Bring your own keys.

Many projects, one window

Set agents going on one project, switch to the next, and come back when they are done. One home screen covers all of them.

Quiet until it matters

Alerts when an agent finishes, goes quiet waiting on you, crashes a run script or hits a merge conflict. Silent for the run you are watching.

Auto-resume

Quit with work in flight. Reopen, click the run, and the conversation picks up where the agent's CLI can continue it.

Race and loop

Send one prompt to four agents and keep the branch you like best. Or loop one agent until your test command exits clean, with hard caps on attempts, time and tokens.

Watch a run, start to finish

An issue handed to an agent, four agents racing one prompt, a loop running until the tests pass, and a branch merged. A one-minute tour, with no sound.

Agency built Agency

Almost every feature on this page was written by agents running in Agency, on worktrees Agency cut, then reviewed and merged in the app.

The commit history is the receipt, and it is public.

Read the commit history

237agent branches merged into main

1,056commits between 18 June and 6 October 2026

Questions

What is Agency?

A desktop app for macOS and Linux that runs AI coding agents in parallel. It launches the agent CLIs you already have, gives each run its own git worktree and branch, and lets you watch, review and merge them in one window.

Is Agency free?

Yes. Agency is free and open source under the Apache-2.0 licence, with no paid tier and no account. You pay only for the agents themselves, through whatever plan or API key each one already uses.

Which coding agents does it work with?

Eleven are built in: Claude Code, Codex, Cursor, Copilot CLI, OpenCode, Gemini CLI, Kimi Code, Crush, Hermes, Pi and DeepSeek Harness. Any other agent you can start from a command line runs too.

How do several agents work on one repo without conflicts?

Each run gets a git worktree: a separate checkout of the same repository, on its own branch, in its own folder. Agents never share a working tree, so they cannot overwrite each other's files. Their work meets only when you merge it, and Agency helps you resolve any conflicts then.

Does Agency send my code anywhere?

Agency does no first-party data collection: no analytics, no telemetry, no account. The agents it launches are third party and send your prompts and code to their own providers, exactly as they do in a terminal. Agency's only request of its own is an update check against GitHub, which you can turn off.

Why not just run agents in tmux?

You can, and plenty of people do, with a worktree per pane. Agency does that bookkeeping for you: it cuts the worktree and branch, tracks which agent is working, waiting or done, shows each run's diff, and merges and cleans up after. The by-hand version is a handful of git and tmux commands, if that is all you need.

Does it run on Windows?

Not yet. Agency runs on macOS 11 or later on Apple Silicon, and on Linux (x86_64 and arm64) as a .deb, .rpm or AppImage. Windows is planned, with no date yet.

Can an agent change things without my say?

Agency never raises an agent's permissions behind your back. Each agent runs with the permission mode you choose for it, in its own worktree, and nothing reaches your base branch until you merge it.

Point it at a repo

Free, with no account and nothing to sign up for. Pre-1.0 and in beta. Windows is planned.

Download for macOS and Linux Requirements