Mindwalk

Mindwalk is a local visualization tool for understanding what coding agents do inside a repository. It replays sessions from Claude Code, Codex, and Pi over a deterministic 3D map of the codebase. Instead of reading a JSONL transcript line by line, reviewers can see where the agent searched, what it read, what it edited, and which files stayed untouched. The README describes the result as a night map: agent activity lights up the parts of the repository it explored while everything else remains dark. The tool is designed for teams that need to review agent behavior, not just final diffs. Session logs answer what happened, but they do not show whether the agent’s path matched the task. Mindwalk adds that missing spatial context. Files can be colored by touch state: seen, read, edited, or unvisited. The UI can show a radial tree or terrain-like map, scrub through a playback deck, reveal compaction or subagent markers, and surface signals such as errors, churned files, and edits after verification. It can also export playback to a WebM file entirely client-side. Mindwalk runs as a Go binary and keeps viewing local. By default it scans common session directories such as `~/.claude/projects`, `~/.codex/sessions`, and `~/.pi/agent/sessions`. It can serve a local UI, open a single session, build a repository map without a session, normalize a trace, or analyze a session with an optional judge through the user’s own Claude or Codex CLI. That optional evaluation sends a summary to the configured model, but normal viewing does not send repository maps or traces to a hosted service. Installation is straightforward for command-line users. The README documents an install script that verifies release checksums and places the binary in `~/.local/bin`, plus a source build path through `make setup && make build`. Windows archives are available from GitHub Releases, and Nix users can install through a listed flake. Mindwalk is best for engineering leads, agent builders, and reviewers who want to see whether an AI coding agent scoped a task well. It can make hidden behavior visible: over-broad exploration, missed relevant files, repeated failed commands, post-test churn, or a tight, well-scoped path. The project is MIT-licensed, so the tool itself is free; any optional LLM-based analysis uses the user’s configured model account. For an OpenTools page, the key distinction is that Mindwalk is observability for agent work, not another coding assistant. It does not replace Claude Code or Codex. It gives teams a local lens for checking whether those agents understood the codebase, stayed inside the expected scope, and left behind a trace that humans can review.