Codewhale: Open-Source Terminal AI Coding Agent

What changed

Codewhale is an open-source AI coding agent that operates within the terminal, built using the Rust programming language. It is designed to read project files, edit code, execute commands, and verify its work, leveraging either hosted or local AI models. The latest release, v0.10.0, introduces a fresh build with continuous community improvement as a core focus. Installation for macOS and Linux is facilitated via a curl script that fetches the official GitHub release, while Windows users can download installers or archives from GitHub Releases. Existing installations can be updated using the codewhale update command. The initial run directs users to a composer interface, requiring a model connection via /provider or F3 to add hosted keys or select local runtimes. It automatically connects to Ollama if it's running with a chat model. Codewhale also offers support for various packaging routes including npm, Cargo, Docker, Nix, Scoop, Android/Termux, and a CNB mirror. Tab completion is available for bash, zsh, fish, powershell, and elvish shells.

Users can initiate tasks by running codewhale in their project directory, selecting a provider and model using /provider and /model commands, and then describing a concrete task, such as "Fix the failing tests and explain what changed." Alternatively, tasks can be executed non-interactively using codewhale exec "task description". The agent supports different operational modes: /mode plan for exploration without file changes or shell execution, and /mode work for making modifications. Users can also specify access levels via Shift+Tab, choosing between Ask, Auto-Review, or Full Access, with detailed explanations available in the modes and permissions guide.

Codewhale includes a terminal interface, a local web client accessible via codewhale web, and a planned native desktop app. A community-maintained VS Code extension also connects to the local Codewhale Runtime. The agent offers features to manage long-running tasks, save sessions, set durable goals, review workflows, and coordinate agents. Safety features include approval modes, repository rules, and optional OS sandboxing to limit agent actions. Users can undo or restore workspace changes using /undo and /restore commands.

Why it matters for builders

Codewhale brings powerful AI coding assistance directly into the developer's terminal, reducing the need to switch contexts. Its ability to integrate with both local and hosted models provides flexibility and control over AI processing and costs. Developers can leverage Codewhale for automated code fixes, task execution, and complex problem-solving directly within their existing development environment.

Practical impact

Developers can install Codewhale using the provided script for macOS/Linux or download installers for Windows. After installation, they can connect their preferred AI models using the /provider and /model commands. For immediate use, running codewhale in a project directory and issuing a task command like "Refactor this function" allows for direct interaction. For scripting or CI/CD pipelines, codewhale exec can be integrated. Builders should explore the different access modes (Ask, Auto-Review, Full Access) to find the right balance of automation and control for their workflows. The VS Code extension offers an alternative integration point for those who prefer an IDE-centric experience.

Caveats and source limits

The source indicates a "fresh release" for v0.10.0, but specific performance benchmarks or detailed comparisons to previous versions are not provided. Pricing for hosted models is not detailed within the source, and users are responsible for managing their own model costs. While the agent supports various local model runtimes like Ollama, vLLM, and SGLang, the setup complexity for these local environments is not elaborated upon. The exact capabilities and limitations of the "Computer Use" plugin, which allows interaction with other applications, are described as requiring review of requested access and platform requirements, with detailed setup guidance available separately. The native desktop app is mentioned as a future direction, with its current availability and feature set not fully detailed in this excerpt.