Claude Code started hiding its steps. This tool reads the logs anyway.

What it does

claude-devtools is a desktop app and optional standalone server that parses the session transcripts and tool logs Claude Code already drops in ~/.claude/. It renders them in a navigable UI: expanded tool calls with inline diffs, per-turn token attribution across seven categories, nested subagent execution trees, and the extended thinking content the terminal suppresses. It also surfaces per-project MEMORY.md layers and can inspect remote sessions over SSH.

The interesting bit

It never wraps or intercepts Claude Code; it simply reads existing JSONL logs, so it requires no API keys and is immune to most upstream changes. The standout feature is a compaction visualization that shows exactly when the context window hits its limit, what gets compressed, and what is discarded—turning a black-box mechanism into a readable timeline.

Key highlights

• Per-turn token attribution across 7 categories (global and project CLAUDE.md, skills, @-mentioned files, tool I/O, thinking, team overhead, user text).
• Recursive subagent trees with per-agent token metrics, duration, and cost.
• A project memory viewer that renders hidden MEMORY.md layers with wikilink navigation and one-click handoff to VS Code, Cursor, Zed, or terminal emulators.
• A standalone server mode that runs without Electron and makes zero outbound network calls, suitable for offline or Docker deployment.
• Notification triggers for sensitive events such as .env file access, tool errors, high token usage, or custom regex patterns.

Verdict

Heavy Claude Code users who have ever stared at a Read 3 files summary and wondered what actually happened will find this essential. If you do not use Claude Code, or you are content with opaque terminal output, there is nothing here for you.