codewhale

Codewhale is an open-source coding agent that runs in the terminal. Written in Rust, it reads a project, edits files, runs commands, and checks its own work using a hosted or local model chosen by the user. The project began as `deepseek-tui` and retains configuration and session compatibility, but is now provider-neutral and independently maintained, with no affiliation to any model provider. Installation on macOS and Linux is via an official install script that selects the latest published release; Windows users download an installer or archive from GitHub Releases. An updater command checks for and applies newer builds. npm and Cargo are supported as secondary packaging routes, alongside Docker, Nix, Scoop, Android/Termux, and an optional CNB mirror. Shell completions are generated per shell. Usage centers on running `codewhale` inside a project folder, selecting a provider with `/provider` and a model with `/model`, then describing a concrete task. Tasks can also be run non-interactively with `codewhale exec`, which suits scripts and CI jobs. The agent can read repositories, edit files, run commands, inspect results, and keep working toward a goal. A plan mode allows exploration without file changes or shell execution, while a work mode permits changes; keyboard shortcuts switch between Ask, Auto-Review, and Full Access permission levels. Beyond the terminal interface, a bundled local web client connects to the same Codewhale Runtime, and graphical web and desktop workbenches are described as in development. A bundled VS Code extension connects to the local Runtime from a sidebar. A Computer Use plugin, included in the source, adds tools for observing and interacting with other applications, subject to OS permissions and platform requirements. Notable capabilities include connecting hosted providers or local models through Ollama, vLLM, or SGLang; inspecting proposed actions and resulting file changes; approval settings with hard policy boundaries even under Full Access; undo and restore for workspace changes; saved sessions, durable goals, workflow review before execution, and coordination of agents with different models and roles; and extension through MCP servers, skills, hooks, and readable agent role files. Unknown model prices are reported as unknown rather than free. Documentation covers providers and local models, agent teams, MCP, hooks, configuration, the local web client, authorization order, modes and permissions, installation, and the repository layout. The project is MIT licensed, with third-party notices recorded, and welcomes bug reports, feature ideas, and pull requests.