Originally published on andrew.ooo — visit the original for any updates, code snippets that aged out, or follow-up posts.

TL;DR

Orca is the open-source "ADE" (Agent Development Environment) for developers running multiple coding agents in parallel. It just hit 7,747 stars with 1,949 new in the last 7 days. The pitch: run Claude Code, Codex, Cursor, OpenCode, or Pi side-by-side, each in its own isolated git worktree, in one desktop app — with a real mobile companion to steer the swarm from your phone.

• 30+ supported CLI agents: Claude Code, Codex, Cursor CLI, Grok, Copilot CLI, OpenCode, Amp, Pi, Devin, Goose, Auggie, Cline, Continue, Charm Crush, Droid, Kilo, Kimi, Mistral Vibe, Qwen Code, Antigravity, MiMo Code… plus "any CLI agent"
• Parallel worktrees — fan one prompt across N agents in N isolated worktrees, compare, merge the winner
• Mobile companion for iOS + Android — push notifications, send follow-ups from anywhere
• Design Mode — click any element in an embedded Chromium browser, the HTML+CSS+screenshot drop straight into the agent's prompt
• SSH worktrees — run agents on a remote box with full file editing, terminals, port forwarding, auto-reconnect
• Native GitHub + Linear — browse PRs, issues, project boards in-app; open a worktree from a task
• Orca CLI — orca worktree create, snapshot, click, fill so agents can drive Orca itself
• Account switcher — track Claude/Codex usage and rate-limit resets, hot-swap accounts without re-logging in
• Bring your own subscription — Orca is free; you pay only for the underlying agents

If you've been running three terminal panes (Claude Code, Codex, Cursor) and cd-ing between worktrees — Orca is the tool you've been building badly in your head.

Quick Reference

What Problem Orca Solves

In 2026 most "10x" developers are running three to five coding agents at once. The workflow has stabilised in the wild: spin off a worktree, send Claude Code at it; spin off another, send Codex; spin off a third, send Cursor; compare three diffs, cherry-pick the best parts, merge.

The tooling for this is terrible. You juggle tmux panes, git worktree add invocations, browser tabs for PR review, a Linear tab, Discord for notifications, and a vague feeling that you don't know which agent has finished. Most "agent IDE" projects either lock you into one agent (Cursor, Windsurf) or assume you run them one at a time.

Orca's bet is that the parallel part is the whole point — you want one orchestration shell for any agent, and a phone in your pocket that pings you when any of them finishes. Built by Stably AI, the project ships daily; the 1,949-stars-this-week jump aligns with the v0.x release that added SSH worktrees and Design Mode.

Why It's Trending NOW (June 2026)

Three things converged this week:

1. The mobile companion shipped to general iOS App Store. TestFlight-only until early June; the full release brought a wave of fence-sitters.

2. Account switcher + usage tracking solved a hot problem. Anthropic and OpenAI tightened rate limits and tier visibility in May/June 2026. Orca's hot-swap between Claude accounts + live rate-limit resets made it the answer for the "how do I survive multiple Claude tiers" complaint.

3. SSH worktrees landed. Cloud-dev workflows (Codespaces, Gitpod) live in a browser. Orca — run agents on a beefy remote box but keep the local desktop polish — picked up the "I don't want to live in a browser" segment.

The Mental Model: ADE vs IDE

Orca's framing is that an ADE (Agent Development Environment) is different from an IDE in three structural ways:

If you've spent any time running agents at this scale, you've felt every one of those gaps. Orca is the first tool I've seen treat them as design primitives rather than features bolted on top of an editor.

Parallel Worktrees — The Headline Feature

The parallel-worktrees flow is the strongest demo of why Orca exists. You write one prompt — say, "Add OAuth2 support to the user service, with tests" — and Orca:

1. Creates N isolated git worktrees under .orca/worktrees/.
2. Launches a different agent in each worktree (Claude Code + Codex + Cursor, or any subset).
3. Streams all N agents' output into a tabbed pane.
4. When all N finish, shows a side-by-side diff comparison.
5. Lets you cherry-pick hunks across agents into a final merge worktree.
6. Notifies your phone when any needs input or finishes.

The novel UX is the merge worktree — instead of three diffs in three terminals and git cherry-pick-ing by hand, Orca gives you a 3-pane diff with checkboxes per hunk. Check the Claude Code result for the auth headers, the Codex result for tests, the Cursor result for the README; Orca builds a merge commit on a fresh worktree.

Supported Agents

Orca treats the agent as a swappable runtime, not a vendor lock-in. The supported list is genuinely "any CLI agent" — they ship pre-configured integrations for the heavyweights:

The bring-your-own path is one TOML config away — if your-agent --version works in a shell, Orca can wrap it.

Real Code: What Driving Orca Looks Like

The Orca CLI makes everything scriptable:

# Create a worktree for a task
orca worktree create feature/oauth-support --agent claude-code

# Fan a prompt across multiple agents
orca worktree fan "Add OAuth2 to user service with tests" \
  --agents claude-code,codex,cursor --count 3

# Snapshot worktree state
orca snapshot create --name "before-merge"

# Drive the embedded browser (Design Mode)
orca browser click "#submit-button"
orca browser fill "input[name=email]" "test@example.com"

# Computer Use — let an agent operate desktop apps
orca computer-use --agent claude-code --task "open Notion and find the Q3 spec"

The config lives at ~/.config/orca/config.toml:

[agents.claude-code]
command = "claude-code"
default_model = "claude-opus-4-7"

[worktree]
base_dir = ".orca/worktrees"
auto_cleanup = false

[ssh.remote-box]
host = "build-server.internal"
user = "andrew"
identity = "~/.ssh/id_ed25519"

The SSH block unlocks "run agents on a 96-core remote box, drive them from a MacBook Air." Auto-reconnect and port forwarding work without intervention; the local app feels identical whether the worktree is at ~/repo or on the other end of SSH.

Design Mode — The Sleeper Feature

The feature most underrated in the README is Design Mode, which is what every web-app dev has been wanting in a coding agent UI for two years.

You open an embedded Chromium browser on your localhost dev server. You click any element on the page. Orca packages up:

• The element's HTML
• The element's computed CSS
• A cropped screenshot of just that element

…and drops the bundle into your agent's prompt textarea pre-formatted. You add "make this card responsive on mobile" and hit send. The agent has everything it needs without you copy-pasting selectors or describing the visual layout.

This is the workflow that paid IDEs (Cursor, Windsurf) have been promising for a while and which Orca shipped first as open source.

Mobile Companion — Is It Useful?

The honest answer: yes, for specific patterns. The mobile app isn't a full agent IDE — you can't write code on your phone. What it does well:

• Push notifications when any agent finishes
• Read the agent's last few messages from the lock screen
• Send a short follow-up ("approve," "rollback," "try a different approach")
• See live usage / rate-limit status for Claude and Codex accounts

In practice: kick off three parallel agents on the desktop, walk away, get pings when each finishes, glance at the diff on the phone, approve or reject. Enormously better than "remember to check back in 45 minutes."

Honest Limitations

1. Daily-ship velocity = occasional rough edges. The README says "we ship daily, so this feature list is perpetually behind." Bugs sometimes ship too; fixes typically land within 24 hours. Pinning a version for team rollouts is wise.

2. Electron-class app footprint. Not Tauri — Orca is a full Electron-style app. ~250 MB DMG, 400–800 MB idle RAM with a few agents running.

3. The mobile app needs the desktop running. Mobile is a companion that pairs with your desktop session over LAN or Stably's relay. Closed laptop = dark phone.

4. N parallel agents = N parallel rate-limit hits. Running Claude Code three times in parallel uses 3x your Anthropic quota. The usage tracker shows it clearly, but new users get caught by surprise.

5. Linux is AppImage only. No .deb/.rpm/Snap/Flatpak yet. AppImage works everywhere but lacks native polish.

6. SSH worktrees require your remote box to be Orca-aware. Not "any SSH server" — Orca installs a small agent on the remote box. One command, documented, but worth noting for locked-down corporate boxes.

Community Reactions

The growth pattern (7.7K stars, 1.9K this week, daily release cadence on the releases page) and active Discord point to real product-market fit among the "I run multiple coding agents" crowd. The Stably team triages feature requests on Discord within a day.

Practical observations from the issue tracker:

• Cross-agent workflows are the most-requested feature. Power users want "Claude Code plans, Codex writes tests, Cursor implements." Orca is moving toward this with worktree-fan, but explicit pipelines aren't in v0.x yet.
• Android lags iOS by about a release.
• Linear/GitHub native integrations are the killer feature for teams. Reviewers consistently call out "this is the first time the project board and the IDE feel like one tool."

How It Compares

The honest read: Cursor and Windsurf are tightly integrated single-agent products with their own AI offerings baked in. They win if you want one polished agent and don't care about parallelism. Orca wins the moment you want two agents, mobile notifications, or vendor-neutral choice — and its bring-your-own-subscription model means you're not paying twice.

Install in 60 Seconds

# macOS (Homebrew)
brew install --cask stablyai/orca/orca

# macOS (Apple Silicon, direct DMG)
open https://github.com/stablyai/orca/releases/latest/download/orca-macos-arm64.dmg

# Windows
start https://github.com/stablyai/orca/releases/latest/download/orca-windows-setup.exe

# Linux (AppImage)
wget https://github.com/stablyai/orca/releases/latest/download/orca-linux.AppImage
chmod +x orca-linux.AppImage
./orca-linux.AppImage

# Arch (AUR)
yay -S stably-orca-bin

First launch walks you through agent detection — it scans $PATH for installed CLIs and pre-populates the agent list. Add your Claude Code / Codex / Cursor accounts via the account switcher, then create your first worktree.

FAQ

Is Orca a replacement for Cursor or Windsurf?

Not directly. Cursor and Windsurf are opinionated single-agent IDEs with their own AI baked in. Orca is the layer above those — it runs Claude Code, Codex, Cursor CLI, and 27 other CLI agents in parallel worktrees. If you only use one agent and like its UI, stay there. If you switch between agents or want parallelism, Orca is the orchestrator.

Do I need to pay for Orca?

No. Orca itself is free and Apache 2.0. You pay for the agents you choose to run (Claude Code via Anthropic, Codex via OpenAI, Cursor CLI via Cursor). The bring-your-own-subscription model means Orca doesn't add a markup or middle-layer billing.

Does the mobile app cost anything?

No. The iOS app is free on the App Store; the Android APK is free on GitHub releases. The companion app requires you to pair with a desktop Orca instance — there's no cloud-only "mobile-first" mode.

Will running 3 agents in parallel burn my Claude / Codex quota 3x faster?

Yes. The account switcher / usage tracker shows live consumption per provider, but the math is what you'd expect: three parallel Claude Code agents = 3x token spend. Many users mitigate this by running cheaper models (Haiku, GPT-5.5 Instant) in parallel and reserving Opus / Codex-Maxxing for the merge agent.

Can I use Orca on a Linux server with no GUI?

Partially. The Orca CLI (orca worktree create, etc.) works headlessly. The desktop UI requires a display. The common pattern is desktop Orca on your laptop + SSH worktree pointing at a headless Linux build box.

How does Design Mode handle JS-heavy SPAs?

It uses the live DOM of the embedded Chromium window, so React, Vue, Svelte, and SolidJS all work — you click the rendered element, you get the live HTML and computed CSS regardless of how it was generated. The cropped screenshot is captured from the rendered Chromium frame, not the source.

Is my code uploaded anywhere?

No. Orca is a local desktop app; your code lives in your local git worktrees and your local agent runs. The mobile companion pairs over your local network when possible, with optional encrypted relay through Stably's servers when you're on different networks (off by default — explicit opt-in).

Verdict

Orca is the first agent IDE that takes "multiple agents in parallel" seriously as the default workflow rather than an advanced feature. The combination of worktree fanout, mobile companion, SSH agents, and Design Mode is exactly the toolkit a developer running Claude Code + Codex + Cursor every day has been hand-rolling badly in tmux.

If you're already paying Cursor or Windsurf and use one agent, stay there. If you've started writing your own scripts to coordinate three agents and a project board — uninstall those scripts and try Orca for a week. The bring-your-own-subscription model and Apache 2.0 license make the downside negligible.

Pair it with Agent Reach for free internet access inside agents and Voicebox for voice I/O, and the open-source agent-developer stack starts to look complete.

Download Orca: github.com/stablyai/orca · onorca.dev