A desktop app that reconstructs exactly what Claude Code did from raw session logs. Every file path, tool call, and token is structured and searchable with no setup required.
claude-devtools is an open-source Claude Code session viewer that transforms the opaque summaries of your AI coding assistant into a fully transparent, searchable record of every action. It reads the raw session logs that Claude Code silently stores in ~/.claude/ and reconstructs the complete picture—every file touched, every tool invoked, every line of reasoning—giving developers, team leads, and security engineers a window into the black box of AI-assisted development. Unlike terminal output that now hides critical details, this desktop app delivers a structured, multi-pane interface where no action goes unseen. Available for macOS, Linux, and Windows, it requires no API keys, no internet connection, and no modification to your Claude Code installation. Its core promise is simple: your Claude is no longer coding blind.
Since Claude Code v2.1.20, the terminal output that developers trusted was deliberately replaced with opaque summaries. Detailed thinking steps, exact tool inputs and outputs, subagent activity, and file-level diffs were all stripped away, leaving users to guess what their AI assistant actually did. This change triggered immediate community backlash because it broke the feedback loop essential for debugging, auditing, and trusting an automated tool that modifies codebases. Without visibility, mistakes go unnoticed, security implications go unchecked, and learning from the assistant’s reasoning becomes impossible. claude-devtools addresses this pain point directly by unlocking the rich telemetry that Claude Code already records but hides. It restores the missing transparency, allowing developers to inspect every decision and every action with the same ease as reviewing a pull request.
Context Reconstruction is the centerpiece feature, offering per-turn token attribution across seven distinct categories: CLAUDE.md files at global, project, and directory levels; skill activations; @-mentioned files; tool I/O payloads; internal thinking tokens; team and subagent overhead; and user text. The interface visualizes exactly how the context window fills during each turn, compacts under memory pressure, and refills afterward—letting you see at a glance what influenced the model at any point. This is not a raw token dump; it’s a curated breakdown that maps every byte to its source, so you can trace a hallucinated suggestion back to the exact @-mention or CLAUDE.md snippet that triggered it. By revealing context dynamics, the tool helps you prune unnecessary content, reduce costs, and improve response quality without guesswork.
The Tool Call Inspector replaces the terminal’s collapsed summaries with a fully expanded, syntax-highlighted view of every interaction. When Claude Code reports it “Read 3 files” or “Edited main.py,” claude-devtools shows the actual content it read, the precise diff that landed, and the stdout or stderr of any Bash command it executed. Read calls are displayed with language-aware highlighting, Edit diffs appear inline with addition and removal markers, and complex Bash pipelines are broken down so you can inspect exit codes and output streams separately. Crucially, subagent execution trees are untangled: each agent gets its own isolated trace with recursive nesting, token counts, duration, and cost. This inspector eliminates the need to pipe `--verbose` output through grep and manually reconstruct the narrative—it presents the full, linear story of the session in a copy-friendly, searchable format.
Publisher
A
admin
Launch Date2026-02-17
Platformdesktop
Pricingfree
Tech Stack
#Docker