{"slug": "orca-115k-stars-for-the-harness-that-runs-your-agents", "title": "Orca: 115k stars for the harness that runs your agents", "summary": "Stably AI's open-source Orca desktop app, which runs coding agents side by side in isolated git worktrees, has reached 115,729 GitHub stars, 7,308 forks, 19,145 commits and 1,000 tagged releases since the repository went up on March 17. The MIT-licensed TypeScript app is free for macOS, Windows and Linux and supports Claude Code, Codex, OpenCode, Copilot, Cursor, Kimi, Qwen Code, Pi, Hermes Agent, DeepSeek Harness, Cline, Continue, Amp and Goose, with SSH worktrees letting a laptop drive agents running on a remote GPU machine. Orca positions itself as an orchestrator rather than an agent, ranking fourth by stars behind OpenCode, Codex and Pi, and runs agents on users' own subscriptions without marking up tokens.", "body_md": "Orca: 115k stars for the harness that runs your agents\n\nOrca is an open source desktop app from Stably AI, a Y Combinator-backed company, that runs coding agents side by side. The repository went up on March 17 and now carries 115,729 stars, 7,308 forks, 19,145 commits and 1,000 tagged releases, the newest published hours before this piece. It is MIT licensed, written in TypeScript, and free to download for macOS, Windows and Linux.\n\nThe feature that earned those stars is the worktree fan-out. You write one prompt, aim it at five agents, and each run gets its own isolated git worktree, so the branches cannot touch each other. You then compare the diffs in one window, annotate the lines you want changed, and merge the winner. Anything that runs in a terminal runs inside Orca, which is why its supported list reads like a directory: Claude Code, Codex, OpenCode, Copilot, Cursor, Kimi, Qwen Code, Pi, Hermes Agent, DeepSeek Harness, Cline, Continue, Amp, Goose and more.\n\n## What the parallel runs actually buy you\n\nFive agents on one prompt is a different way of buying quality than picking the best model. You are not asking which agent writes the best checkout fix in the abstract. You are asking which one gets your checkout fixed tonight, with your tests, in your repo, and then merging the diff you can read. The comparison happens on output rather than on a leaderboard.\n\nThat shape only works if the agents are cheap enough to run five at once, which connects to the other story of the week. Teams are routing subagents to open models, and a harness that fans work out to five runners makes the cheap tier more valuable, not less. The expensive model does not have to be the one that touches every branch.\n\n## The stack it absorbs\n\nThe README lists what comes in the box, and the list is a stack collapse rather than a feature set: a Ghostty-class terminal with WebGL rendering and scrollback that survives restarts, a VS Code style editor with drag and drop file prompts, worktree management, GitHub and Linear panes, a design mode that captures the HTML, CSS and a cropped screenshot of any element in a real Chromium window and drops it into a prompt, diff annotation that ships comments back to the agent, and a CLI that lets an agent drive Orca itself through worktree create, snapshot, click and fill.\n\nThe commercial logic is in the README’s one line about subscriptions: run any agent with your own subscription. Orca is a client, not a reseller. It does not mark up tokens, which means it also has no reason to steer you toward the model that pays it the most. There is an account switcher that tracks Claude and Codex usage and rate limit resets and lets you hot-swap accounts without logging in again.\n\n## The part that matters for a rig\n\nSSH worktrees are the feature this site should care about. You can point Orca at a remote machine and run the agents there with full file editing, git and terminals, with automatic reconnect and port forwarding. The laptop in your bag holds the window; the machine with the memory and the GPUs does the work. That is the same split every build on this site has been arguing for, arriving in the tool people already have open.\n\nIt also changes what a cheap laptop means. The $1,499 value pick does not have to run the agent session, only the interface to it, and the $5k desk build or the Spark cluster becomes the thing you dial into from anywhere. The hardware question stops being “what fits on my desk” and becomes “what do I leave running at home.”\n\n## Where it sits in the harness war\n\nAgainst the twelve agents already listed on this site’s harness directory, Orca arrives as an orchestrator rather than another agent: 115,729 stars put it fourth behind OpenCode, Codex and Pi, and unlike those four it is not trying to be the agent at all. That is the bet. GitHub shipped a Copilot app with projects and canvases, Microsoft is packaging its harness through Foundry, and Orca is the independent layer that assumes you will keep several agents and want them managed rather than replaced.\n\nThe caveats are the usual ones for a project moving this fast: 4,400 open issues and 3,800 open pull requests, a release cadence measured in hours, and a marketing line about shipping 100x that no benchmark supports. The worktree isolation and the agent-agnostic design are real, and both are visible in the code rather than only in the pitch.", "url": "https://wpnews.pro/news/orca-115k-stars-for-the-harness-that-runs-your-agents", "canonical_source": "https://tokenstead.ai/guides/orca-agent-fleet", "published_at": "2026-11-05 06:20:39+00:00", "updated_at": "2026-11-05 06:21:49.516169+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools", "ai-products"], "entities": ["Orca", "Stably AI", "Y Combinator", "Claude Code", "Codex", "OpenCode", "GitHub", "Microsoft"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/orca-115k-stars-for-the-harness-that-runs-your-agents", "markdown": "https://wpnews.pro/news/orca-115k-stars-for-the-harness-that-runs-your-agents.md", "text": "https://wpnews.pro/news/orca-115k-stars-for-the-harness-that-runs-your-agents.txt", "jsonld": "https://wpnews.pro/news/orca-115k-stars-for-the-harness-that-runs-your-agents.jsonld"}}