What is mindwalk?

Mindwalk is a visualization tool that replays a coding AI agent's session as glowing light moving across a 3D map of your codebase. When an AI coding assistant works on a project, it produces a raw log of what it searched, read, and edited, but reading that log line by line does not show whether the agent explored the right parts of the repository or stayed within the scope you intended. Mindwalk turns that log into a night map of your repo where the files the agent actually touched light up, so the shape of its attention becomes something you can see at a glance instead of read. It is a single Go binary that reads session logs from Claude Code and Codex, runs entirely on your own machine, and does not send anything anywhere just from viewing a session. Installing it is a one line curl script, or building from source with make. Running it with no arguments scans your local Claude and Codex session folders, starts a local web server, and opens a browser showing the visualization automatically. Files are colored by how deeply the agent touched them, from unvisited and dark, through seen, read, to edited, and a playback deck lets you scrub through the session over time with search and read actions shown in cool colors and edits shown in warm colors. The tool also includes an optional evaluation feature where it asks your own locally installed claude or codex command line tool to judge how well a session went, covering things like whether the agent explored efficiently, stayed in scope, or wandered. This is the one feature that sends data off your machine, and only when you explicitly trigger it: a summary of the session, including message wording, file paths, and short event descriptions, is sent to the model behind your own account. Reports from these evaluations are cached locally so they are not regenerated automatically. Under the hood, the project separates the session log, the repository layout, and the evaluation report into three distinct pieces, joined by a local Go server that serves a React and Three.js frontend for the 3D view.