Biscuit

Biscuit's pitch is reduction: fewer megabytes, fewer dependencies, faster startup. The implementation is credible — a Python/Tkinter native app rather than another Electron shell, under 20 MB on disk, launched with a single 'biscuit path/to/src' command after 'pip install biscuit-editor'. That matters most for people working inside VMs, on older laptops, or in environments where installing Node.js alongside an editor is friction they'd rather skip. The agent is the second half of the story. Biscuit documents 11 tools the agent can call: ReadFileTool, EditFileTool, DeleteFileTool, ListDirTool, GlobFileSearchTool, GrepTool, CodebaseSearchTool, RunTerminalCmdTool, TodoWriteTool, GetWorkspaceInfoTool and GetActiveEditorTool. That set covers reading, editing and deleting files, directory listing, glob and grep search, whole-codebase search, running terminal commands, writing a dynamic task list, and reading your workspace or active editor state. In practice that is what a refactor loop needs — find the code, change it, run the tests, keep a task list. The agent streams responses with thought visualization, and it will plan against a dynamic task list rather than answering in one shot. Supported models are named explicitly: Claude 4.5 in opus, sonnet and haiku variants, and Gemini 2.5 in flash and pro. You supply the API keys, which means the editor itself doesn't bundle model costs. There is also a small interaction worth knowing: typing '# your prompt' in the integrated terminal invokes an LLM call, which you then accept or decline. Day-to-day editing is not an afterthought. LSP integration covers 550+ languages for completions, definitions, references, hover docs and a symbol outline. Search is ripgrep-based with regex and bulk replace — the same engine many terminal-first developers already trust. Git is built in: clone, branch, stage, commit, push, pull, a diff viewer, and GitHub issues and pull requests without leaving the window. Python developers get a debugger with breakpoints, variable inspection, call stack and runtime variable modification. Multi-cursor editing and a Vim mode round out the expected conveniences. Extensions are written in Python, and the API can register language servers, formatters, AI providers and UI views. That is a real advantage if you already write Python — you can extend your editor in the same language you work in — and a real limitation if you don't, because the JS/TS plugin ecosystem that VS Code enjoys simply isn't there. Two honest caveats: the documentation is flagged as a work-in-progress, and the Ollama extension for running local models is deprecated, with a rewrite listed as pending, so offline AI is not something to count on today. Where it fits: solo developers, tinkerers, open-source contributors, anyone on constrained hardware or inside a VM who wants agentic coding without the weight. Where it doesn't: teams that need real-time collaborative editing out of the box, anyone whose workflow is defined