claude-devtools 🤖 Agent Open source

What it does

claude-devtools is a desktop and web viewer for Claude Code sessions. It reads the transcripts and logs Claude Code already writes under ~/.claude/ and rebuilds them into a navigable interface, so you can see which files were read, what patterns were searched, the exact diffs applied, subagent activity and how the context window filled up. It does not wrap or modify Claude Code, and it works on past sessions as well as current ones.

What is inside

The main views cover per-turn token attribution (CLAUDE.md files, skills, mentioned files, tool input and output, thinking, user text), compaction visualisation that shows when context was compressed, and a tool call inspector with syntax-highlighted reads, inline edit diffs and Bash output. Subagents and agent teams appear as nested execution trees with their own tool traces, duration and cost. Other features include a project memory browser, cross-session search from a command palette, side-by-side panes, clean copy and export to Markdown, JSON or plain text, SSH access to sessions on remote machines, and notification rules for events such as .env access or high token usage.

Works with

Claude Code only, since it parses Claude Code's own session format. Builds are offered for macOS, Linux and Windows, plus a Docker image that serves the UI in a browser.

How to install

On macOS with Homebrew:

brew install --cask claude-devtools

Or run the standalone server with Docker:

docker compose up

Maintenance and safety

MIT licensed with regular releases and a changelog. The standalone server makes no outbound network calls according to the project, and the README shows how to mount ~/.claude read-only and disable networking entirely. File reads are restricted to the project root and the Claude data folder. Windows and unsigned macOS builds may need a manual trust step on first launch. Session logs can contain secrets, so treat any remote or shared deployment carefully.

Who should use it

Claude Code users who want to debug why an agent did something, audit token usage, understand context loss after compaction, or review subagent runs without reading raw JSON logs.