Lathe

Lathe is an open-source experiment by Deven Jarvis in using LLMs to teach rather than think for you. It pairs a set of coding-agent skills with a Go CLI that stores generated tutorials and serves them in a local reader. The latest release listed is v0.5.0.

What It Is

Lathe generates hands-on, single-part or multi-part technical tutorials from any prompt, such as /lathe build a 3D Slicer in Erlang, run inside a coding agent. The README lists Claude Code, Cursor, Codex, Gemini CLI, opencode, Cline, and Windsurf as supported agents. You then read and work through the tutorial by hand in a local UI started with lathe serve, which serves at localhost:4242 by default. The binary never calls a model itself; all model work runs in your interactive agent session.

How the Workflow Runs

Skills cover the tutorial lifecycle: /lathe writes the first part, /lathe-extend adds the next part, /lathe-verify works through the steps to confirm they compile and run, /lathe-ask answers questions about a part, /lathe-tag adds search tags, and /lathe-voice creates custom voices. With /lathe-work running, the Ask, Verify, and Add a part buttons in the UI drive the agent session directly. Without it, the buttons give you the command to paste.

Reading, Library, and Provenance

The reader provides a table of contents, side-notes, and exercises at the end of each tutorial. Tutorials are stored in ~/.lathe/tutorials/ with a metadata.json recording tags, tool versions, sources, voice, and the model used. The list page offers client-side search, sorting, and filters by status, type, tag, and version. Each tutorial records the URLs consulted and shows an authorship byline.

Writing Voices and Verification

Voices change how the prose sounds without altering structure or accuracy. The plainspoken voice is the default and companion is a warmer first-person option. Verification is opt-in: it runs each step in a scratch directory and records verified, failed, or skipped, with skipped used when the required toolchain is missing. The author states that tutorials are LLM-generated and can contain errors, and that testing so far has been on Claude Code and macOS.

Setup Path

Install via Homebrew cask (macOS), an install script, go install (Go 1.25+), or from source, then run lathe skills install to place skills for your agent. Local models work by pointing your coding agent at an OpenAI-compatible endpoint.