Orca

Give each task an isolated worktree for parallel agents

Orca is worktree-native: each task receives its own git worktree checkout, branch, files, and agent terminals. The same prompt can be fanned out to several agents while their changes remain isolated, making it possible to compare real code diffs rather than just text answers. The workflow follows a consistent Create, Work, Review, Ship, and Archive/Delete lifecycle, with commit, push, pull-request creation, and check status available inline.

Sources: [1][2]

Keep terminal, editor, browser, and diff review in the same worktree

Each worktree can hold agent terminals, editor tabs, browser tabs, and terminal splits. Annotate AI Diffs lets reviewers leave comments on changed lines and send that feedback back to the agent without moving through a separate IDE review loop. Design Mode goes further for frontend work: clicking an element in a real Chromium window can attach its HTML, CSS, and a cropped screenshot directly to the agent prompt, combining DOM context with the visual symptom that needs to be fixed.

Sources: [1][3]

Run long-lived agents remotely and supervise them from desktop, mobile, or CLI

SSH worktrees create the worktree and run agents, git, and terminals on a remote host while keeping editing and diff workflows integrated locally. Orca supports reconnecting, restoring terminal sessions, and forwarding remote ports, which makes it practical to move heavy builds or GPU-dependent work off a laptop. The Mobile Companion focuses on monitoring and steering running agents, including completion notifications and follow-up messages, while the Orca CLI exposes worktree creation, snapshots, and UI automation for scripted workflows.

Sources: [1][4]

For competing implementations and parallel development lanes

Orca is most useful when the same feature or bug fix should be attempted by several agents and only the strongest implementation should be kept, or when frontend, backend, tests, and other streams need to progress concurrently. GitHub and Linear task integration makes it easier to fit that parallelism into an existing Git-centered workflow rather than building a separate AI task system. SSH execution is also valuable when the machine that owns the repository should not be the machine that performs the expensive work. If the workflow is still one agent in one checkout, a normal terminal or IDE may be simpler.

Sources: [1][2][4]

Dependencies, secrets, remote tooling, and agent accounts still need setup

A new git worktree starts as a clean checkout, so gitignored dependencies, caches, and secrets such as node_modules or .env files are not automatically present. Orca provides shared paths, worktree.sharedDirectories in orca.yaml, and .worktreeinclude to share or copy selected paths, but teams still need to decide which local state should be shared and which should remain isolated. More parallel agents also mean more disk use, dependency setup, compute consumption, model cost, and rate-limit pressure, so parallelism does not translate directly into linear speedup.

Orca orchestrates terminal-based agents rather than replacing their runtimes. Installation, authentication, subscriptions or API usage, and permission models remain specific to Codex, Claude Code, OpenCode, or whichever CLI is used. SSH adds host-key, credential, network, and remote-toolchain responsibilities; some Linux remote terminal setups need a C/C++ toolchain and Python for native modules. Orca itself is built as an Electron desktop application for macOS, Windows, and Linux. The repository is licensed under MIT.

Sources: [2][4][5][6]