What is mindwalk?
mindwalk reads the session logs that coding agents leave on disk and plays them back on a map of the codebase. Files glow as the agent searches, reads and edits them, so you can see at a glance where it explored, how far its changes spread and whether it checked its work. A timeline marks context compactions, subagent launches and user turns, and each subagent's trace can be replayed on its own.
It also has an optional evaluation step: your own `claude` or `codex` CLI drafts success criteria from your request and scores the session on four fixed dimensions (exploration, scope, wandering, verification), with every finding linked to timeline events.
It is a single Go binary with a React and Three.js front end, MIT-licensed, by Ricko Yu, the author of the Go live-reload tool Air. v0.5.0 shipped on 7 August 2026, the last commit landed on 10 August, and the repo had about 1,400 stars at last check.
Is mindwalk worth it?
mindwalk answers a question raw JSONL logs do not: what the agent treated as relevant. Each file keeps its deepest touch (seen, read or edited), deleted files linger as wireframes, and the playback histogram colours observation cool and mutation warm, so an agent that edited before it read is obvious at once. A review strip folds error rate, churned files and edits after the last verification into one line, which is the part a reviewer will use daily.
The privacy design is careful. Viewing is fully local and sends nothing; the installer checks the binary against published checksums. Evaluation is opt-in and runs your own CLI with no tools, no MCP servers and no settings, sending only that session's user messages, file paths and one-line event digests. Verdicts are rolled up mechanically from finding severities rather than left to the model, and criteria are reused when the task wording has not changed, so repeat scores stay comparable.
Coverage is the constraint. It reads Claude Code, Codex and pi logs; OpenCode, Crush and Antigravity support exists only as requests and open pull requests, including a nine-part Crush series opened on 9 September and not merged at last check. With 17 open issues and pull requests, five releases between 11 July and 7 August and nothing merged since 10 August, it is a working v0.5 rather than a maintained product.
It sits next to ccusage rather than replacing it: ccusage tells you what a session cost, mindwalk shows what it did. Neither has a team view, and mindwalk only sees sessions still on disk, so Claude Code's default 30-day transcript cleanup limits how far back it can go.
Where it earns its place
A tech lead reviewing pull requests opened by Claude Code
Check whether the agent's changes stayed within the scope of the ticket
Open the session with `mindwalk open <session.jsonl>` and scrub the timeline; amber outside the expected directories is the cue to read the diff closely. Free, and takes a minute per session.
A developer tuning CLAUDE.md or a skill
See whether new instructions change how much the agent explores before editing
Run `mindwalk analyze` on sessions before and after the change and compare the exploration and verification dimensions. Each analysis is up to two calls on your own Claude or Codex plan.
An engineering manager running an agent pilot with five developers
Show the team concrete examples of good and poor agent sessions
Export a playback to .webm from the deck menu, recorded client-side, and share it in a retro. There is no shared server, so each developer exports from their own machine.
Before you commit
- Setup
- `curl -fsSL https://raw.githubusercontent.com/cosmtrek/mindwalk/master/scripts/install.sh | sh` installs to ~/.local/bin after a checksum check; Windows archives are on GitHub Releases. Running `mindwalk` scans ~/.claude/projects, ~/.codex/sessions and ~/.pi/agent/sessions and opens a browser.
- Integrations
- Session adapters for Claude Code, Codex and pi; the `claude` or `codex` CLI as evaluation judge; JSON exports of the trace, the repository map and reports, with schemas in schema/. Nix users can install it via the numtide/llm-agents flake.
- Data and privacy
- Viewing is local and sends nothing. Evaluation, only when you run it, sends one session's user messages, file paths and event digests to the model behind your own CLI; reports are cached in ~/.mindwalk/reports. MIT licence.
- Cost at scale
- Free; evaluation costs up to two model calls per session on your existing plan. No hosted or paid edition.
- Limits
- Three agent formats only; no merged changes since August 2026; serves on a random local port per machine with no multi-user view; history is limited to logs the agent still keeps.
What's good
- Fully local viewing; nothing sent unless you evaluate
- Subagent traces replayable on the same map
- Evidence-linked evaluation with mechanical verdicts
- Single Go binary with checksum-verified installer
What's not
- Only Claude Code, Codex and pi logs today
- No merged changes since 10 August 2026
- No team or multi-machine view
- Evaluation costs tokens on your own plan
Best for
Developers and tech leads who review coding-agent work and want to see an agent's footprint before reading its diff.
mindwalk alternatives
mindwalk: common questions
- Is mindwalk worth it?
- Seeing where an agent wandered in your repository takes seconds with mindwalk, but it is a young tool whose development has gone quiet since August. We rate it 3.6 out of 5.
- How much does mindwalk cost?
- Free and open source (MIT). Session evaluation uses your existing Claude or Codex plan.
- What are the downsides of mindwalk?
- Only Claude Code, Codex and pi logs today; no merged changes since 10 August 2026; no team or multi-machine view; evaluation costs tokens on your own plan.
- Who is mindwalk best for?
- Developers and tech leads who review coding-agent work and want to see an agent's footprint before reading its diff.
- How much does mindwalk cost for a team or in production?
- Free; evaluation costs up to two model calls per session on your existing plan. No hosted or paid edition.
- Is mindwalk safe for company data?
- Viewing is local and sends nothing. Evaluation, only when you run it, sends one session's user messages, file paths and event digests to the model behind your own CLI; reports are cached in ~/.mindwalk/reports. MIT licence.
- What are the best mindwalk alternatives?
- The closest alternatives we have reviewed are ccusage, claude-mem and Graphify.
Answers checked Oct 26, 2026.