Codewhale

Overview

Codewhale is an open-source coding agent and terminal client built on a local Rust agent engine. It can read a project, edit files, run commands, and check its work, while using hosted provider APIs or local and self-hosted models.

Key Features

• Provider choice across more than 40 built-in provider routes, OpenAI-compatible endpoints, and local models through Ollama, vLLM, or SGLang.
• Plan, Work, and Operate modes with approval postures, workspace recovery through /undo and /restore, and session activity listed by /receipts.
• Interactive terminal, headless exec, local browser client, and a local HTTP runtime API for threads, events, and approvals.
• Extensibility through MCP servers, skills, plugins, hooks, and compatibility with existing Claude Code plugins.
• Multi-agent workflows with sub-agents, supervised agent teams, Fleet setup and status, and checked-in workflows.
• Computer Use plugin for observing and operating other applications, with access granted by the user.

Use Cases

• Build an app, add a feature, or write a script, then run the code and test what it builds.
• Automate repeated workflows from the terminal, scripts, or CI, including headless turns with codewhale exec.
• Connect tools through plugins and MCP servers, or use APIs from your own scripts.
• Split a larger job across agents with different roles and models using Fleet.

Getting Started

• Install Codewhale on macOS or Linux with the install script, follow any PATH instruction it prints, then open a project folder.
• Connect a model with codewhale auth set --provider deepseek, use /provider or F3, and switch models with /model.
• Run codewhale and describe a concrete task, then choose permissions and review its work.
• Optionally save a multi-agent team with /fleet setup and check runs with /fleet status.

Deployment & Requirements

• Runs on macOS, Linux, and Windows; Android on Termux is a preview.
• Install routes include the install script, npm, Cargo, Docker, Nix, Homebrew on Linux, and manual downloads.
• Model access can use provider accounts, an OpenAI-compatible endpoint, or local and self-hosted inference through Ollama, vLLM, or SGLang.
• Connected services require their own setup and authentication.
• The terminal, local browser, and VS Code GUI need no Codewhale account; hosted web and desktop use an account.

Before You Adopt

• License: MIT. Review its terms before using, modifying, or distributing the project.
• The hosted web app is a development preview, and hosted task execution is still being qualified.
• The native desktop app is in development and has no public download yet.
• Connected services require separate setup and authentication.
• On Linux, bubblewrap sandboxing depends on bubblewrap being installed and working.