Stably AI's Orca turns the agent IDE into a fleet command center

Stably AI released Orca as an open-source desktop and mobile application that lets developers run multiple coding agents in parallel, each in its own isolated git worktree, and compare the results side-by-side. The pitch is direct: fan one prompt across five agents, watch them compete, merge the winner. Orca is an ADE — agent development environment — built for the reality that no single coding agent is consistently best.

The tool supports any CLI-based agent: Claude Code, Codex, Grok, Cursor, GitHub Copilot, OpenCode, Devin, Cline, and a dozen others. Each agent gets its own terminal, its own worktree, and its own full file system context. Orca renders terminals with Ghostty-class WebGL rendering, infinite splits, and scrollback that survives restarts. It ships with a mobile companion for iOS and Android that notifies users when an agent finishes and lets them send follow-ups from their phone.

What makes Orca notable is not the feature list. It is what the tool assumes about the state of AI coding in mid-2026.

The bet is that agent diversity beats agent fidelity.

Every major coding agent lab has spent the last two years trying to build the single best model for code generation. Anthropic pushes Claude Code. OpenAI pushes Codex. GitHub embeds Copilot into its IDE. Each vendor wants developers to commit to one agent, one subscription, one workflow. Orca treats that assumption as obsolete.

Instead of asking which agent is best, Orca asks which agent is best for this specific task. The user writes one prompt. Orca fans it across five agents in parallel. The agents produce five different solutions in five different worktrees. The developer reviews the diffs, picks the winner, and merges. The losing worktrees get discarded.

This is a fundamentally different mental model from the single-agent IDE. It is closer to ensemble methods in machine learning than to traditional software engineering. The developer becomes a judge and a merger, not a writer. The agents do the writing. The human does the selection.

Orca’s design reflects a specific insight about current agent capabilities. Coding agents are good but inconsistent. Claude Code might nail a refactor that Codex fumbles. Codex might generate a test suite that Claude Code overcomplicates. Grok might produce a novel approach that neither considered. Running them in parallel and picking the best result is a cheap way to compensate for individual agent weaknesses. It is a practical hedge against the fact that no frontier model has solved code generation.

The parallel worktree feature is the core mechanism. Orca creates a fresh git worktree for each agent, so each agent operates on its own copy of the codebase. The developer can compare diffs across worktrees, annotate specific diff lines with comments, and ship those comments back to the agent for revision. The review loop stays inside Orca. The developer never leaves the tool.

Orca also ships a Design Mode that lets users click any UI element in a real Chromium window to send its HTML, CSS, and a cropped screenshot into the agent’s prompt. This is a narrow but powerful capability for frontend work. Instead of describing a UI bug in prose, the developer points at it. The agent gets exact visual context.

The mobile companion is the most surprising feature. It suggests that Orca’s designers expect agents to run for minutes or hours, not seconds. A developer can kick off a parallel agent run, walk away, and get a push notification when the agents finish. They can review the results and send follow-ups from their phone. This is a workflow designed for asynchronous, long-running agent tasks, not for real-time pair programming.

Orca is licensed under MIT and is free and open source. Stably AI ships daily updates, and the GitHub changelog is the real feature list. The project has attracted a community on Discord and Twitter, with the @orca_build account pushing regular updates.

What Orca means for AI builders.

The tool signals a shift in how developers think about coding agents. The single-agent IDE — Cursor, Windsurf, Zed with Claude — assumed that one model, one agent, one workflow would be sufficient. Orca assumes the opposite. It assumes that developers will run multiple agents, from multiple vendors, in parallel, and treat the ensemble as the reliable unit.

This has implications for the agent economy. If developers run five agents per task, they need five subscriptions. Orca supports account switching and usage tracking, letting users see Claude and Codex usage and rate-limit resets, and hot-swap accounts without re-logging. The tool does not abstract away the subscription cost. It makes it visible and manageable.

For agent vendors, Orca is both an opportunity and a threat. It increases the surface area for agent usage — developers who would not commit to one agent might run five. But it also commoditizes the agent layer. If the developer is selecting the best result from five agents, the individual agent brand matters less. The ensemble is the product. The developer’s loyalty is to Orca, not to Claude Code or Codex.

The open-source license matters. Orca is not a proprietary platform that locks developers into a specific agent ecosystem. It is a shell that runs any CLI agent. Developers can swap agents in and out, run experimental agents alongside production agents, and build custom workflows without vendor approval. This is the Unix philosophy applied to AI coding: small, composable tools that do one thing well.

Orca’s parallel worktree approach also highlights a limitation of current agents. If one agent were consistently the best, there would be no need to run five. The fact that Orca exists and has traction is evidence that agent quality is uneven and task-dependent. The ensemble method is a workaround for a problem that the agent labs have not solved.

The tool is early. The GitHub repository is active but young. The mobile companion is in beta. The feature list is long but unevenly polished. What matters is the direction.

Orca treats the developer as a manager of agent fleets, not as a writer of code. That is the most honest reflection yet of where AI coding is in 2026.