Image: stablyai/orca
Why it mattersOne workbench for every coding agent means a team can test which agent does better on the same ticket without juggling four CLIs, four subscriptions and four worktrees.
Picking one coding agent used to mean picking a lane. Orca, released as open source by Stably AI, is a workbench that runs 50 of them side by side on your own API keys, and its GitHub page shows 101,690 stars with 6,400 forks on 22 October.
Orca lists the agents it supports on its README: Claude Code, Codex, Grok, Cursor, GitHub Copilot, Muse, DeepSeek Harness, OpenCode, Pi, Cline, Continue, Devin and about 40 others. Trendshift ranked it 16th on its daily list on 21 October. The licence is MIT and the project is marked yc-backed in the repository topics.
What a reader can actually do with it
Orca opens one task across multiple git worktrees at once, so a developer can hand the same ticket to Claude Code, Codex and Grok and compare the three diffs when they return. The README calls the agent outputs annotatable, which means a reviewer can mark up the diff in the Orca UI before accepting any one of them. The project ships a desktop app, a mobile companion for iOS and Android to watch runs from a phone, and SSH-based remote worktrees for jobs that need a bigger machine than a laptop.
The subscription model does the saving
Orca is bring your own key: the user supplies their Anthropic, OpenAI, xAI and other provider credentials, and Orca talks to each provider directly. There is no middle fee on model calls. That matters because the main cost of running five coding agents is five subscriptions, not five binaries, and reusing one subscription per vendor across tasks is where the saving sits. The project does not state a price for Orca itself, so the whole bill is whatever the chosen agents cost.
What this does not tell you
The repository page does not state when Stably AI founded the project or when the first commit landed, so the 15,693 commit count is a total rather than a rate. Nothing in the README benchmarks one agent against another on code quality or cost, which is a judgement a team still has to run on its own code.
For a team already paying for two coding agents, this is one place to see which one wins on a given task without opening two shells. For a team paying for none, the time to start is a different decision, and this item is not a reason to make it.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.