tomerzipori/local-agent-toolkit

Local Agent Toolkit is a command line tool that lets Codex, Claude Code, or a human developer hand off small, well defined coding tasks to a language model running locally through Ollama. The idea is to save the more expensive cloud based AI tokens for the hard decisions, while a local model on your own computer handles routine work like exploring a codebase, doing a first pass review, brainstorming test ideas, running diagnostics, or drafting a candidate patch. The local model is treated as a helper, not a final authority. Anything it produces is considered untrusted advice that a human or the parent AI agent still needs to check before accepting it. Commands cover several stages of a coding workflow: picking a suitable installed model, exploring code with find and files, planning a change, reviewing staged or branch level work, drafting tests, diagnosing failures, getting a second opinion on a design, and writing a small patch. Most commands only read information rather than changing anything. The patch and write tests commands print out suggested changes without applying them automatically, and only one command, fix test, will actually run a shell command, and only the exact command a person supplies and approves. By default, the toolkit only looks at files tracked by git, skips ignored, untracked, sensitive, binary, or oversized files, and limits how much file content and how many files get sent to the model. There are opt in flags for cases where broader access is genuinely needed, such as including untracked files or allowing a remote host instead of the default local one. Setup requires macOS with zsh as the primary supported environment, Python 3.10 through 3.13, Ollama with at least one installed model, and Git. Ubuntu is tested in continuous integration, though Windows and non zsh setups are not yet supported. After installing, the toolkit can add a matching skill file for Codex or Claude Code so those tools know how to call it. Configuration, including the Ollama host, context window size, and file size limits, is stored in a local config file and can be adjusted from the command line. The project is early stage and welcomes contributions such as new commands and support for more platforms.