Codewhale is a coding agent that reads a project, edits files, runs commands, and checks its own work from the terminal, using a hosted or local model you choose. It began as deepseek-tui, still reads that project's configuration and sessions, and is now provider-neutral.
Every client drives the same local Codewhale Engine, so sessions, tools, and permissions behave the same across the interactive terminal interface, headless one-turn runs from scripts or CI, a bundled browser client, an advisory pull request review command, and a local HTTP Runtime API for threads, events, and approvals. A native desktop app is in development, and a community VS Code extension connects to the same Engine.
Work is split into modes and approval postures. Plan mode explores without changing anything, Work edits and runs, and Operate drives a goal through planned, verified steps. Postures (Ask, Auto-Review, Full Access) decide when a tool call needs approval, and commands run inside an OS sandbox where supported (Seatbelt on macOS, bubblewrap on Linux).
For longer jobs, Codewhale sets durable goals, delegates bounded work to sub-agents, and runs supervised agent teams (a Fleet, with a different model per agent) that can be scripted as checked-in workflows. It extends through MCP servers, skills, plugins, and hooks, and loads existing Claude Code plugins.
Features
- Model choice: over 40 built-in provider routes (Anthropic, DeepSeek, Google, Mistral, Moonshot, OpenAI, OpenRouter, xAI), any OpenAI-compatible endpoint, and local models through Ollama, vLLM, or SGLang
- Account keys: an optional Codewhale account stores provider keys for use across signed-in devices, while local use works without an account
- Modes and postures: Plan, Work, and Operate modes combined with Ask, Auto-Review, and Full Access approval postures
- Recovery and receipts: undo and restore commands recover workspace changes, and receipts list every file, command, and approval in a session
- Long-running work: durable goals, sub-agents, supervised agent teams with a pre-spend check, and checked-in workflows
- Extensions: MCP servers, skills, plugins, hooks on session and tool events, and compatibility with Claude Code plugins
- Computer Use plugin: an included plugin adds tools for observing and operating other applications, enabled after review
- Request preview: shows the exact redacted request before anything is sent to a model
- Pull request review: an advisory review of a pull request, with posting opt-in
