Hmbown/CodeWhale

Open-source, community-driven agent harness

What it solves

Codewhale is a terminal-based coding agent that allows developers to automate software engineering tasks directly from the command line. It removes the dependency on a single AI model provider by allowing users to "bring your own model," bringing together various hosted and local LLMs into a single, unified toolset for editing files and running commands.

How it works

The agent operates by reading your code, editing files, and executing shell commands to complete a given task. It can be used interactively via a Terminal User Interface (TUI) or headlessly via the codewhale exec command for scripts and CI/CD pipelines. It supports a wide range of providers (including DeepSeek, Claude, GPT, and local options like Ollama or vLLM) and allows users to switch models mid-task using the /model command.

Who it’s for

Developers who want a model-agnostic coding agent that runs locally on their machine and provides granular control over permissions and security sandboxing.

Highlights

  • Model Agnostic: Supports 30+ providers, including hosted services and local LLMs.
  • Security-First: Includes a read-only "Plan" mode, approval gates for risky commands, and OS-level sandboxing (Seatbelt on macOS, bubblewrap on Linux).
  • Resumable Work: Uses an append-only ledger to record every step, allowing users to resume tasks via fleet resume.
  • Multi-Interface: Offers a TUI, a headless execution mode, and a local browser client.
  • Constitution-Based Control: Supports a constitution.json file to define write holds that cannot be bypassed.