Stably AI Launches Orca: An Agent Development Environment for Running Fleets of Parallel Agents

Software development startup Stably AI has launched Orca, an open-source Agent Development Environment (ADE) designed to run and orchestrate multiple artificial intelligence coding assistants simultaneously. First released on GitHub in March 2026 and widely discussed across developer platforms by May 2026, the platform acts as a desktop control plane that allows developers to run command-line interface (CLI) agents—such as Claude Code, Codex, and Grok—side-by-side.

Written in TypeScript and licensed under the MIT License, Orca is available as a desktop application for macOS, Windows, and Linux. The platform introduces a dedicated interface built specifically to handle the operations of autonomous agents rather than capturing direct human keystrokes like a traditional integrated development environment (IDE).

---

Isolating Parallel Agents in Git Worktrees

The primary engineering problem Orca seeks to address is the state collision that occurs when multiple AI agents attempt to operate on the same repository at the same time. When multiple autonomous agents are deployed simultaneously within a standard single-threaded codebase workspace, they typically trigger file system conflicts, locked files, and race conditions as they modify overlapping files.

Orca resolves this state collision by using isolated Git worktrees. A Git worktree allows a developer to have multiple working directories checked out from a single local Git repository. Under this architecture:

• Hermetic Sandboxing: Orca spawns a hermetically sealed sandbox directory (worktree) for each individual agent.
• Independent Execution: Agents can modify, compile, and test code independently within their designated worktrees.
• Clean Working Directories: The main working directory remains clean and unaffected while the agents complete their tasks.
• Developer Control: Once the agents finish, the human developer can review the diffs, annotate code lines, and merge the selected solution into the primary branch.

This design allows a developer to “fan out” a single prompt to multiple competing agents. For instance, a developer can assign a single bug to Claude Code, Codex, and Grok concurrently to compare their generated solutions side-by-side in real time.

---

Local Hardware Management Through Agent Hibernation

Executing multiple active CLI agents in parallel can create significant hardware overhead, as each running agent functions as an active pseudo-terminal session consuming system memory and CPU cycles. To prevent local hardware exhaustion on desktop machines, Orca implements a mechanism called agent hibernation.

When a developer switches focus away from a particular worktree tab, the system suspends the idle terminal and serializes its application state to the local disk. This process frees up system RAM while preserving a complete snapshot of the agent’s work, enabling near-instantaneous resumption when the developer selects the tab again.

Additionally, for resource-heavy operations such as compiling massive codebases, Orca supports remote execution through SSH worktrees. Instead of checking out branches locally, the environment establishes a secure shell connection to a remote server, creates the Git worktree on the remote file system, and routes the terminal’s input and output back to the desktop client.

These remote sessions run via a persistent relay process on the host machine. If the local desktop client loses its network connection, the remote agent continues its execution uninterrupted until the developer reconnects.

---

Scripting Workflows via Headless RPC Server

On an architectural level, Orca utilizes an internal Remote Procedure Call (RPC) server that communicates via WebSockets. This design exposes the development environment’s core state machine to programmatic manipulation.

Using the Orca CLI, developer tools and external scripts can programmatically drive the IDE itself. This enables headless, automated workflows such as:

# Example of using Orca to run a CLI agent in a dedicated worktree
orca worktree create --agent claude-code --prompt "Fix failing tests in checkout.ts"

In a continuous integration (CI) pipeline, for example, a test failure runner can programmatically call the Orca CLI to spin up a background worktree, inject the failing stack trace into a specific command-line agent, and assign the debugging task entirely without human intervention.

---

Native Integrations and Licensing

Orca functions as an orchestrator and does not bundle proprietary model weights or subscription packages. The environment operates on a “bring your own subscription” model; users connect their pre-existing API keys or terminal-authenticated accounts for the CLI agents they wish to run.

The platform features built-in support for more than 30 CLI-based agents, including Claude Code, Codex, Grok, Gemini, Cline, Cursor, OpenCode, and Pi. It also includes native integrations with GitHub and Linear for tracking issues and pull requests, alongside a mobile companion app designed to let developers monitor and prompt active agents remotely. The mobile companion app is available for iOS via the App Store and TestFlight, and for Android via a direct APK download from GitHub.

Stably AI, the company behind the project, was founded in 2022 by Jinjing Liang, a former senior software engineer on the Google Chrome team, and Neil Parker, a former Uber tech lead. The team previously launched Stably, an AI-powered end-to-end software testing platform designed to generate and maintain Playwright tests in CI pipelines.