Why Orca Passed 80,000 GitHub StarsOne Place to Run, Review, and Return to Parallel Agent Work

1. The appeal: keep your favorite agents, and see the work clearly

Once asking AI to write code becomes routine, another problem appears. A terminal is fixing a bug. Another session is building a feature. A browser holds the preview, and a diff waits for review. All of them matter. As they multiply, it becomes harder to answer a simple question: which job needs me now?

Orca is a development environment for that situation. It runs agents such as Claude Code and Codex and brings their workspaces, terminals, browsers, and reviews together. Its developers call it an ADE, or Agent Development Environment. The product organizes work with existing agents rather than supplying a new AI model.^[1]

On September 30, 2026, its official GitHub repository displayed 81.6k stars; the API returned 81,628. For a repository barely six months old, that is a striking number. Stars measure interest rather than usage or performance, but what drew that interest is worth examining.^[2]

The central argument of this article is that Orca combines the freedom to use different CLIs with less effort spent organizing and checking their work. Preparing workspaces, commenting on diffs, watching usage, and knowing which job needs a reply: work that people once handled by stitching together terminals, editors, and browsers is taken on by one product.

For someone whose work already fits comfortably inside one official app, with little parallel activity, the benefit is modest. To find the dividing line, this article follows a day of work.

2. How the attention grew: coverage and user reports point to not losing track of work

GitHub records the repository's creation date as March 17, 2026. A repository creation date does not necessarily establish a public launch or general availability date. It does show how young this project is relative to the attention it has attracted.^[2]

Reading coverage in chronological order provides some context. A July Writeble introduction foregrounded running multiple CLIs in separate worktrees. Its references to both a July launch and a March initial release are a reason to avoid assigning a single definitive birthday from that article.^[3]

An August 23 Reddit report was more specific about day-to-day value. Its author described organizing sessions across several repositories and two computers, with useful visibility into their states. Other participants preferred environments they had assembled themselves. These are individual experiences, but the emphasis is revealing: the praise is not that the AI became smarter, but that the author stopped losing track of work.^[4]

A September 4 hands-on article from Tagbangers traced the author's path through tmux, cmux, and herdr before trying Orca, including having different agents produce alternative versions of a screen. A September 19 AIgent Lab introduction offers a more recent entry point into installation and parallel work.^[5][6]

The product's design suggests three sources of appeal. People can try it without abandoning CLIs they already know and pay for. Its parallel workflow is visually easy to demonstrate, which suits video and social media. And a free, MIT-licensed application lowers the barrier to trying it.

Many of the firsthand reports come from people who had built their own terminal setups with tools such as tmux. Receiving that self-assembled workshop as a product appears to be what first resonated, which also matches the themes of the early coverage.

3. Why build around CLIs? There is more to choice than choosing a model

A CLI is an interface operated through commands in a terminal. Launch Claude Code in an Orca terminal and you are fundamentally using Claude Code. Launch Codex and you are using Codex. Their subscriptions and usage limits do not become one shared pool just because their windows share a home.

The useful distinction is between a model and an agent. Even with the same model, an agent's file search, tool use, permission handling, conversation storage, skills, and configuration can differ. An app that offers several models and an environment that runs several agents provide different kinds of choice.

Using CLIs as the foundation lets developers retain each agent's own machinery while standardizing the workspace around it. A new agent can enter through the terminal as well. But basic terminal compatibility and deeper integration—status detection, usage tracking, or chat rendering—are separate levels of support. Check the features you need for each agent.^[7]

Orca's documentation identifies its terminal as xterm.js-based. It offers themes, shortcuts, splits, and scrollback search, although the experience need not match a favorite standalone terminal exactly. The design preserves terminal workflows while adding graphical tools around them.^[8]

Display options are expanding too. An experimental Chat UI sits alongside an updated structured chat surface for eligible local Claude and Codex sessions. Existing terminal sessions, SSH, WSL, and other environments can follow different paths, and the feature is still being refined.^[9]

For a CLI enthusiast, this is a shared workshop that accommodates familiar tools. For someone who prefers chat, it is another interface to weigh against the polish of the official apps. The product is widening to serve both entry points.

4. Beyond splits: connect setup, review, and requests for changes

Imagine working on one web service: fixing search, adjusting a pricing page, and adding tests. This is an illustrative workflow, not a measured speed comparison.

Give each job a workspace

Git worktrees provide separate working directories for the same repository. Orca treats these as units of work and associates terminals and views with them. Returning to a job becomes easier when its branch and changes have an identifiable place.^[10]

Opening another terminal and creating another worktree are different operations. Two terminals in the same worktree still access the same files. Before letting agents edit concurrently, check which workspace each session actually uses.

Worktrees help prevent the confusion caused by simultaneous edits to the same physical files. If separate branches change the same function, conflicts can still appear at integration. Even a clean textual merge can combine incompatible assumptions about behavior. Server ports, shared databases, and external services also need their own arrangements.

Read the diff and reply at the relevant line

Annotate AI Diff lets you attach comments to changed lines and send them to an agent as a batch. One comment might identify a missing signed-out case; another might request a boundary test. The target and feedback travel together, reducing repeated copying of code fragments and line numbers.^[11]

What gets shorter is the loop of finding a change, understanding it, requesting a revision, and checking the result. The more code AI writes, the more the human reading-and-replying step determines overall speed. Orca invests heavily in that step.

Keep usage limits in sight

A quieter convenience lives at the bottom of the window. For supported accounts, the status bar shows usage and time until reset for applicable windows, including five-hour and weekly limits. Settings let you choose percentages used or remaining. These describe allowance consumption rather than a monetary bill.^[26]

Before starting another long task, you glance down. With room available, you continue; with a tight limit, you might review existing work first. Opening settings or switching to a browser kept aside for usage checks becomes less necessary. Small interruptions disappear. That kind of placement can shape everyday comfort more than a list of headline features.

Freshness depends on the underlying agent’s usage records, so treat the display as a planning aid rather than a second-by-second meter. The information you want while working sits exactly where your eyes can reach it.^[26]

Point at the visual problem

Design Mode lets you select an element in the built-in browser and send its surrounding HTML, computed CSS, and a cropped screenshot to the agent. Source location information can also be included when suitable development metadata is available.^[12]

Instead of describing “the little button at the top right that needs more padding,” you select the button and explain the intended change. For frontend developers and designers, this connects visual feedback with concrete context. Mapping a selected element back to its source is not always straightforward, so checking the revision remains essential.

5. Alongside the official apps: vendor development interfaces and where Orca fits

By September 2026, vendors' own development apps also offer worktrees, parallel sessions, and visual review. Listed feature by feature, the gaps are small; the differences lie in what each product puts at the center.

This is a starting point for comparison, not a complete feature matrix or performance ranking. Availability depends on plan and execution environment. Each row is grounded in the vendor's documentation.^{[13][14][15][16][17][18]}

OpenAI's June 23 guide already describes starting, reviewing, and organizing work from a phone, including worktrees and inline review. Continuing agent work away from the desk is becoming part of competing products too.^[13]

Claude Desktop's Code interface offers multiple sessions, Git isolation, and previews. A coherent official interface may be particularly attractive to someone less comfortable with terminals. Google describes Antigravity 2.0 as a standalone agent-management application, moving the focus toward assigning work and inspecting results.^[14][15]

Cursor is an especially important comparison: its April 2 Cursor 3 announcement included Agents Window and Design Mode. Pointing at interface elements for feedback and isolating work with worktrees are ideas several products have already adopted.^[17]

Devin's official examples include launching three sessions to explore competing solutions. Its axis is how much work to delegate and where it should execute, a different way of handing off work from Orca's arrangement of CLIs on your own machine.^[18]

Seen side by side, Orca's distinguishing choice is to center the design on bringing each vendor's CLI, configuration, and subscription as they are. Vendor apps refine their own agent experience; Orca provides a workshop that stays the same when the agent changes. For people who have assembled similar setups from a terminal, an editor, a Git GUI, and virtual desktops, the question is whether to hand off building and maintaining that arrangement.

6. How Agent Teams relate: a system for collaboration and a system for organizing workspaces

Claude Code Agent Teams coordinate multiple Claude Code sessions through shared work and direct messaging. A lead can divide tasks and synthesize results. Official documentation describes the feature as experimental and disabled by default.^[19]

Being able to ask a team to divide three issues and coordinate is powerful. The same documentation also advises avoiding concurrent edits to the same file. Even with a team that can talk, deciding boundaries and dependencies up front is the most direct way to prevent conflicts.^[19]

If the search behavior has not been agreed on, implementing its tests in parallel can create rework. It may be better to settle the specification first, then separate implementation and testing. That is a question of task design before it is a question of which application to install.

Orca terminals placed side by side do not consult each other. Instead, Orca provides an experimental Orchestration layer with tasks, dependencies, supervised workers, messages, and decision gates. Its scope extends from arranging work for a human to providing machinery for agents to assign and track it.^[20]

Agent Teams is a substantial option for collaboration among Claude sessions. Orca organizes an environment that can include different agents and optionally their coordination. The two can coexist: Claude Code can run inside Orca. In either case, finding genuinely independent work matters more than increasing the number of workers.

7. Screen space and human attention: jobs running versus jobs being read

Place long conversations, code diffs, and a browser into four panes on a 13- or 14-inch laptop and each window becomes too small for detailed diffs. The right number of splits depends on text size, resolution, and the task.

But the number of jobs running need not equal the number of jobs being read. While the search fix runs tests, you can review the pricing page. Orca exposes agent states, and its experimental Agent Dashboard groups work that needs attention, is running, or has finished. You no longer have to watch every log; you return when called.^[21]

Multiple windows and virtual desktops

On a larger display, Orca's documented pane system can combine terminals, diffs, and browsers, saving the layout per worktree. Its Agent Dashboard can also open in a separate pop-out window.^[22][21]

Opening several main workspaces and keeping each worktree on its own macOS Space, and restoring that arrangement after a restart, could not be confirmed from official documentation. If that setup is essential to your workflow, verify it on your own machine before switching.

SSH connections and self-hosted Orca Servers concern where work executes. Virtual desktops concern where windows appear. Moving execution to another machine can reduce local compute load, but it does not expand the reviewer's attention.^[23]

People who have carefully arranged terminal windows and virtual desktops may work more happily in their existing setup. Orca helps most not by showing more at once, but by reducing the search when you return to a job you left.

8. The people behind it, and the business behind free software

Orca is built by Stably AI. Y Combinator lists it as a Winter 2022 company and names Jinjing Liang and Neil Parker as founders. Their profiles describe Liang as a former senior Google Chrome engineer and Parker as a former Uber Tech Lead. The company page also describes Stably's AI testing product, giving the team a background in development and software testing.^[24]

Orca itself is free under the MIT license. Agent subscriptions, API usage, and rented servers remain separate costs. A self-hosted Orca Server also uses compute you provide; it is not a promise of free hosted infrastructure.^[2][23]

An Enterprise page offers direct discussions about deployment, approved providers, organization defaults, and support. That is a visible commercial entry point.^[25]

The structure suggests free open-source adoption leading to enterprise support and related business. Orca's revenue mix, path to profitability, and future paid plans are not visible in public materials.

For organizations, the decision also involves actual data flows and management controls. The Enterprise page's “SOC 2 Readiness” wording should not be treated as verification of an audit report. The OSS license, application operations, and AI-provider contracts are separate considerations.^[25]

Sending the same bug to three agents produces more candidates, along with more processing and comparison work. Parallelism may reduce elapsed waiting without reducing total usage. Dividing independent work and racing several solutions to the same problem are also different uses of concurrency.

9. Try a complete work cycle, not a bigger wall of terminals

Orca looks especially relevant to someone who uses several CLIs, progresses independent tasks in the same repository, and personally reviews the resulting diffs. That pattern matters more than whether the developer works alone or for an enterprise. A solo developer can have three independent fixes in flight; a large-company engineer can be perfectly comfortable completing one task at a time in an official app.

A small trial can make the choice concrete. The following is a proposed evaluation method, not a vendor benchmark.

1. Choose two small, independent changes. Initially avoid tasks that both edit the same central files. Give each a one-sentence completion criterion.
2. Compare the whole cycle in Orca and your usual environment. Include setup, waiting, diff review, follow-up instructions, and integration. Model response speed alone misses much of the product's purpose.
3. Record three kinds of effort. Time spent finding the next job that needs you, actions needed to send review feedback, and integration rework. Also observe usage and local resource load.
4. Use your normal display arrangement. A laptop alone, an external monitor, and virtual desktops create different constraints. If multiple main windows are essential, check both their behavior and restoration first.
5. Keep the simpler arrangement if it wins. If the work does not become easier, continuing with your current app or terminal environment is a valid outcome.

For further reading, follow the July introduction, the August user discussion, the September 4 firsthand account in Japanese, and the September 19 introduction. Finish with the current official documentation to separate earlier expectations from today's features.

For a video entry point, Michael Sahlmann Diaz has a Spanish-language stream titled “Probando Orca ADE: Cómo programar con MÚLTIPLES Agentes de IA en paralelo | 054”. For a quick feel of the interface, the short clips in the official site's feature demos are a good start.

Orca is a response to a practical question: when several AIs are working, where does the person return to understand and check the results? Individual features exist elsewhere. Still, keeping how you survey work and reply to it consistent while changing CLIs has value of its own. If that makes your day easier, it is a far better reason to adopt Orca than any star count.

Research scope: product analysis based on official documentation, public source code, and firsthand user reports. Long-term use, speed comparisons, and multi-Space operation were not measured. Illustrations are AI-generated conceptual images, not product screenshots.

References

1. [1]Orca: What is Orca?. Intended users and the bring-your-own-agent design. ↩
2. [2]stablyai/orca and GitHub API. Retrieved September 30, 2026: 81,628 stars; created March 17, 2026; MIT license. Counts change. ↩
3. [3]Writeble, July 13, 2026. Historical coverage, not authority for current capabilities. ↩
4. [4]Reddit: Orca ADE is incredible, August 23, 2026. Individual experience and discussion. ↩
5. [5]Tagbangers hands-on account, September 4, 2026. Japanese-language firsthand testing. ↩
6. [6]AIgent Lab, September 19, 2026. Introductory reading; current specifications checked against official material. ↩
7. [7]Orca: Supported agents. Terminal compatibility and agent-specific integrations. ↩
8. [8]Orca: Terminal. xterm.js, customization, and state visibility. ↩
9. [9]Orca: Chat UI. Experimental surfaces and environment differences. ↩
10. [10]Orca: Worktrees and Git: git-worktree. Separate working directories. ↩
11. [11]Orca: Annotate AI Diff. Also checked in the public repository's documentation source. ↩
12. [12]Orca: Design Mode. Element context, CSS, screenshots, and available source locations. ↩
13. [13]OpenAI: Mastering remote engineering work from your phone, June 23, 2026. Worktrees, review, and remote work. ↩
14. [14]Anthropic: Claude Code Desktop. Parallel sessions, Git isolation, review, and previews. ↩
15. [15]Google: Antigravity 2.0 Overview. Standalone app and agent management. ↩
16. [16]Google: Settings Overview. Local and Worktree project settings. ↩
17. [17]Cursor: New Cursor Interface, April 2, 2026. Agents Window, parallel work, and Design Mode. ↩
18. [18]Cognition: Fix Checkout Latency with Three Competing Strategies. Parallel-session example and usage trade-offs. ↩
19. [19]Anthropic: Agent Teams. Experimental status, shared tasks, messaging, and avoiding conflicts. ↩
20. [20]Orca: Orchestration. Experimental tasks, workers, messages, and decision gates. ↩
21. [21]Orca: Agents & sessions. Agent Dashboard and pop-out window; repository source also checked. ↩
22. [22]Orca: Tabs, panes & split layouts. Splits and per-worktree layouts. ↩
23. [23]Orca: Ways to run Orca and Remote Orca Servers. Execution locations and self-hosting. ↩
24. [24]Y Combinator: Stably AI (Orca). Founder biographies, Winter 2022, and Stably's testing business. ↩
25. [25]Orca: Enterprise. Deployment discussions and public operating policies; not revenue or profitability evidence. ↩
26. [26]Orca: Usage & rate-limit tracking. Time windows, reset information, used/remaining display, and data freshness. ↩