Orca ADE: parallel CLI agents, each in its own git worktree
As of 2026-10-19 the official repository calls Orca an AI Orchestrator for 100x builders, and the docs site calls it a desktop IDE: it runs command-line agents such as Codex, ClaudeCode, OpenCode or Pi side by side, each in its own git worktree, while model access still comes from your own subscription. This page covers what parallelism actually saves, what it costs you, and how Orca pre-fills each CLI's permission flags.
Updated 2026-10-20
Point this tool at the QCode endpoint — one key for Claude, GPT and Chinese models.
- Billed per token — live rates on /models
- Pay by card (Visa / Mastercard), Apple Pay, Google Pay or crypto
- Self-service: top up and activate instantly
Four things to know
Official name: ADE, also AI Orchestrator
The official README now opens with Orca as an AI Orchestrator for 100x builders, while the docs site still calls it a desktop IDE. Both point at the same thing: it orchestrates command-line agents you already run, and no model is bundled into it. The site page footers still call it the agent development environment (ADE). Checked 2026-10-19.
Its own checkout per task
The README's Parallel Worktrees entry fans one prompt out to several agents, each in its own isolated git worktree, then compares the results and merges the winner. The docs add the mechanism: every task gets its own on-disk copy of the repo through git worktree, so agents never step on each other's files. A brand-new worktree is a clean checkout, so dependencies, caches and local secrets that live in gitignored paths have to be recreated.
The combobox just starts a process
The supported-agents page puts the mechanism plainly: the agent combobox just launches a process in a terminal, so any CLI agent that runs in a terminal can be used. The built-in picker ships a one-click list; deeper hooks, status, usage tracking and account switching are documented per agent. The README now shows Codex, ClaudeCode, OpenCode or Pi running side by side, each in its own worktree, tracked in one place.
The usage bar reads local files
The official usage-tracking page says Orca reads the usage state each agent keeps on its own disk and shows it in the status bar, with one line right after it: no extra API calls, no extra auth. The consequence follows: the numbers update when the agent writes them, not in real time. Clicking that segment opens the Usage popover with every tracked provider, its plan, the soonest reset and the per-window bars.
What it is, and what it is not
Official sources describe the same product in three ways: the repository README calls Orca an AI Orchestrator for 100x builders, the docs index calls it a desktop IDE, and the site footer still calls it the agent development environment (ADE). All three mean the same thing — several command-line agents, each in its own git worktree, running in parallel with the progress visible in one place. What it is not matters more: the official docs state plainly that it is not a model, models and subscriptions stay on your side, and it does not sell managed compute, so the machines and cloud accounts you use remotely remain under your control (README and docs checked 2026-10-19).
The numbers this half year
No hype numbers, only values readable on the page that day. The data embedded in the GitHub repository page puts createdAt at 2026-03-17; the release carrying the Latest badge is v1.4.230. The README's first screen says free and open source under the MIT License, so installing it, reading it and filing issues costs nothing up front; the visible line on the site reads Free and open source. New features shipped daily. Spending starts on the agent side. Behind all the talk sit two things the vendor's own pages back: open source under the MIT License, and a cadence the vendor itself calls daily, with the README list perpetually behind. These values drift, so they stay here only as a 2026-10-19 snapshot.
Timeline
2026-03-17: the repository page on GitHub lists createdAt as this very day (UTC); the same metadata says it is not a fork, is owned by an organisation and is not archived. As of 2026-10-19, the repository had been public for less than a year.
2026-09-20: the latest release at the time was v1.4.206, and its notes say a landed PR usually takes 48–72 hours to be released (except P0+ fixes), while the README admits it ships daily, so that list is perpetually behind. The install docs set the scope: auto-updates follow the stable channel by default, stable releases are vetted, and new features show up first on the RC channel, often daily. Checked 2026-10-19.
2026-10-19: the latest official release is v1.4.230. The GitHub Releases page lists it as published 2026-10-19T06:30:00Z with draft and prerelease both false, so it is a regular release.
Verbatim in the docs vs traps
Verbatim in the docs
Checked 2026-10-19. Verbatim on official pages: the model side, the official wording is that Orca runs agents you already use and the subscription sits on your side; the compute side, it does not sell managed VPS hosting; the isolation side, Orca is worktree-native, so each task gets its own on-disk copy of the repo rather than branching and stashing; the fit side, it works with any CLI agent, since the combobox just launches a process in a terminal — though picker entries are preconfigured for one-click launch, with deeper hooks, status, usage tracking and account switching noted only where supported. The usage bar reads local state each agent keeps on disk, with no extra API calls and no extra auth, and numbers update when the agent writes, not in real time. Install is one command, brew install --cask stablyai/orca/orca, and on Linux the CLI lands as orca-ide. License: free and open source under the MIT License. On remote work, a release note says Orca Server 2.0 is coming soon to make remote work more reliable, while the current Remote Server remains experimental and may have bugs. On release rhythm, the install docs say stable releases are vetted and RC builds ship new features first, often daily.
⚠️ What official sources leave open, and the name clash
The name clash has a second layer: Plotly's chart export tool is also called orca, and its README says it has been deprecated as of August 2026 and points at Kaleido; the bare orca cask in Homebrew describes that same tool (interactive Plotly charts), and it has been disabled since 2026-09-01. Three things official sources leave open. First, the release note about Orca Server 2.0 says it is coming soon without a date, and the Remote Server you can use today is still marked experimental and may have bugs. Second, the README admits its feature list is perpetually behind the releases and points to the changelog as the real list, so command names, package names and agent names are only valid as of the day they were checked. Third, the name clash: on GNOME desktops /usr/bin/orca belongs to the screen reader, which is why the official docs install the Linux command line as orca-ide and explicitly say not to verify with command -v orca; spell the Homebrew command out as stablyai/orca/orca instead of the short orca.
Against N terminals plus tmux
What it adds
Four things the docs state and a bare terminal makes you build: a separate on-disk repo copy per task; one prompt fanned out to several agents, then compared and the winner merged; a usage bar reading local bookkeeping instead of asking the vendor; and a mobile companion that pushes a notification and takes follow-ups. The official README also says: Ghostty-class terminals with WebGL rendering, infinite splits and scrollback that survives restarts; in Design Mode, clicking any UI element in a real Chromium window sends its HTML, CSS and a cropped screenshot straight into your agent's prompt. Checked 2026-10-19.
What you give up
What you give back: a new worktree is a clean checkout, and the docs say plainly that dependencies, caches and local secrets in gitignored paths stay missing until you recreate them; agents launched in parallel come with each CLI's permission-bypass flag pre-filled; the command-line orchestration has to be enabled in the settings first; the mobile side is read-mostly, and the desktop stays the single source of truth. If you open your own terminals and tmux instead, all of that stays in your hands — scripts stay readable, reviewable and versionable — but the isolation layer and the usage view are yours to build.
How to install it, and the first run
① macOS: the Homebrew cask uses the name the docs give, stablyai/orca/orca, with the command brew install --cask stablyai/orca/orca; the cask tracks the stable channel, and brew upgrade --cask orca picks up new stable builds. Untapped, brew install --cask orca hits Plotly's chart exporter, disabled in Homebrew since 2026-09-01, so it fails instead of installing the wrong app. RC builds come from the GitHub Releases links or the in-app Check for Updates. ② Windows and Linux: the install page gives download entries for macOS, Windows and Linux; the AppImage, .deb and .rpm packages for Linux contain the same app and differ only in how updates reach you — only the AppImage lets Orca update itself, while .deb and .rpm just tell you that a new version exists and hand you the install command. ③ First launch: the docs say the app is signed and notarized, and macOS may still ask you to confirm on first launch, which is normal for Electron-based apps. ④ Command line: register the CLI under Settings → General → Orca CLI; on Linux the command is orca-ide, not orca, because /usr/bin/orca belongs to the GNOME screen reader, and the docs explicitly say not to verify with command -v orca. ⑤ After installing, create a worktree and run a small task first, and remember that a brand-new worktree is a clean checkout: dependencies, caches and local secrets that live in gitignored paths are not there.
Using Orca with QCode
Orca neither proxies models nor keeps credentials for you: the first-session page says that a new worktree launches the CLI you picked in the correct working directory and forwards your existing subscription credentials, so endpoint and key changes go into that CLI's own configuration — and that key can be a QCode key. That is the only bridge between the two. Model calls are billed per token, and the per-model prices are listed on /models; Orca itself is a local app, not a model offered here. To point Claude Code at a different base_url, see /anthropic-base-url-setup-guide; the endpoint fields per agent are on /orca-third-party-api-key. If you want agents to run for long stretches away from your desk, or a fixed remote development environment, QCode cloud servers fit: each one has 1 dedicated public IPv4 and full root rights, and you connect with ssh root@IP after provisioning, with IP replaced by the address from the panel. Terms are 1, 3, 6 or 12 months, and there is a 1-day trial billed by configuration; if provisioning fails, or the server is not delivered within 2 hours of the order, the fee is refunded to your balance in full; with auto-renewal on, one month is renewed from your balance 3 days before it expires, and the panel shows the status and lets you start, stop, restart, renew and reset the password; the system image ships without a site-building panel, and right now you cannot switch lines or change the configuration online. For starting prices, see the line comparison on qcode.cc/en/cloud.
FAQ
Do I need a Claude or Codex subscription before Orca?
Not the app itself. Orca is open source under the MIT License, and the README states that it does not sell model access: the models come from the subscription you already pay for, which the docs describe as bringing your own Claude, Codex or OpenCode subscription. The bill belongs to the agent, not to Orca — the docs spell it out that you already pay for Claude Code, Codex or Cursor CLI, and the GLM page says outright that Orca does not include or resell GLM access. Checked 2026-10-19.
Can I point it at my own key or endpoint?
You do not paste them into Orca. The first-session page says that a new worktree opens an agent combobox and that Orca then launches the chosen CLI with the correct working directory and forwards your existing subscription credentials; the supported page puts it plainly: the agent combobox just launches a process in a terminal. Endpoint and key changes go into that CLI's own configuration. The one model switch on the Orca side is --model in orchestration, which the docs limit to opaque provider model IDs for Claude, Codex and Cursor; every other agent goes back to its own configuration. The field-by-field write-up lives on /orca-third-party-api-key. Checked 2026-10-19.
Why doesn't its usage bar show OpenRouter or another gateway?
Because the usage bar reads the bookkeeping each agent keeps on its own disk rather than a live vendor endpoint. The official usage-tracking page names the local state of Claude Code, Codex, Gemini, OpenCode, Kimi Code and MiniMax, and concludes with one line: no extra API calls, no extra auth — so the numbers only move when the agent writes them. Whether the bar stays accurate after you point an agent at a third-party endpoint such as OpenRouter is not covered by the official docs.
How far does the phone app go?
You can watch and send follow-ups, but it is not a second client. The README says you can monitor and steer your agents from your phone; the mobile page pins the scope down: a read-mostly view — agent status, recent terminal scrollback, plus replying to a prompt, sleeping a worktree, reviewing source control and switching agent accounts. Pairing is one-time and the desktop is always the source of truth, and the companion is still in beta. Codex rate-limit reset credits can be spent from the companion's Accounts screen. Checked 2026-10-19.
An SSH worktree refuses to connect — what first?
The official ssh page names two things that get in the way. ① Host key verification: Orca checks built-in SSH connections against your effective OpenSSH known_hosts files and the keys it saved before; if the host was rebuilt and its fingerprint changed, the connection is rejected before you are asked for a password, so verify the new fingerprint yourself first, and for a known_hosts mismatch Orca hands you the matching ssh-keygen -R command. ② The remote toolchain: on first connect Orca installs a small relay on the remote machine, and on Linux that often means compiling node-pty on the spot; if the remote host lacks make, a C++ compiler and python3, Orca still finishes connecting for files, git and the editor, but remote terminals will not work until the build tools are installed.
Does it call models or rack up bills on its own?
Not on the usage side: the official text says Orca only reads the bookkeeping each agent keeps on its own disk, with no extra API calls and no extra auth. What actually calls models and produces the bill is the CLI you started. The same docs warn about two things: for new launches Orca pre-fills each supported CLI's permission-bypass flag, and a worktree is an isolated checkout, not a security sandbox. The documented countermeasure is to choose Manual in Settings → Agents → Agent Permissions. Orca adds no model proxy layer, since the supported page says the agent combobox just launches a process in a terminal. Checked 2026-10-19.
Sources
Official repository and README: https://github.com/stablyai/orca and https://raw.githubusercontent.com/stablyai/orca/main/README.md. Official docs: https://www.onorca.dev/docs — the pages quoted word for word on this page are install, model/worktrees, ssh, mobile, agents/supported, agents/usage-tracking, cli/orchestration and first-session (raw sources under https://raw.githubusercontent.com/stablyai/orca/main/docs/site/content/docs/). Releases and release notes: https://api.github.com/repos/stablyai/orca/releases. Command names, package names and agent names are only valid as of the day they were fetched.
Get the key and the quota straight first
Orca is a local app: it ships no model, and the model calls come from the configuration of the agent you start, whose key can be a QCode key. Not ready for a monthly plan? Pay-as-you-go Packs start at $4 for a one-time quota you use within 7 days (top up from $5).
Related reading
Claude Code custom endpoint setup
Pointing that agent at another base_url: the two variables live there.
Orca third-party endpoints and keys table
The endpoint fields and the official wording for each agent are listed there.
Claude Code Worktree parallel workflow
Running parallel worktrees by hand, without Orca, is covered there.
Crypto payment guide
Nine coins, 56 networks, zero platform fee for AI coding API credits.
This page paraphrases the publicly available Orca documentation and repository from Stably. It is not affiliated with Orca or Stably and carries no endorsement from them. Command names, package names, agent names and the behaviour of the usage bar are limited to what the official pages showed on 2026-10-19; the README itself says its feature list is perpetually behind the releases, so check the original pages again before you install. QCode is only one possible provider of the key your agent uses and is not responsible for how Orca behaves.