Claude DevTools
The missing debugging tool for Claude Code — visualize hidden session logs, tool calls, token usage, and context window in a clean UI.
Evidence shows: app reads local ~/.claude logs, no outbound network calls (except auto-update and SSH), strict IPC input validation, path containment checks, sensitive credential paths rejected. But user confirmation mechanism is weak, auto-update and SSH connections lack explicit user confirmation. Data flow transparency is documented but not code-level evidence. Sensitive data handling has path restrictions but no detail on handling sensitive info in logs. Dependency security only listed in package.json, no vulnerability scan or audit evidence. External effects include auto-update and SSH, but documented. Rollback mechanism not mentioned. Source attribution clear but publisher unverified.
Evidence shows: code structure consistent, tests cover IPC validation and global search, but no detailed failure message examples. Dependency availability reflected in package.json and CI config, but actual availability not verified. Failure messages not fully demonstrated in tests.
Evidence shows: targets Claude Code users, multiple installation methods, multi-platform and Docker support. Capability boundaries clear (not a wrapper, read-only logs). Trigger precision in notification feature with custom regex, but not detailed. Environment fit good, supports macOS/Linux/Windows/Docker.
Evidence shows: clear information architecture, detailed README, complete installation instructions. Naming stability not explicit, version 0.1.0. Few examples and FAQ, known limitations not explicitly listed. MIT license clear, changelog exists. Maintenance responsibility via GitHub Actions and community channels.
Evidence shows: high output usability, multiple export formats and copy features. High marginal value, solves Claude Code log visualization problem. Cost-benefit reasonable, free open source, zero configuration.
Evidence shows: README claims partially supported by docs, but no code-level verification. Cross-source corroboration limited, only README and tests. Fact-inference separation unclear, some feature descriptions may be inferred.
- Auto-update and SSH features may introduce external network interactions; users should be aware.
- Dependency security lacks vulnerability scanning or audit evidence; consider checking dependency versions.
- Publisher identity unverified; assess source credibility cautiously.
What does this agent do, and when should you use it?
claude-devtools is an open-source desktop application (built with Electron) for Claude Code that reads the session logs already stored in ~/.claude/ on your machine and reconstructs the detailed activity the terminal collapses into terse summaries like "Read 3 files". It provides a tool call inspector with syntax-highlighted file contents and inline diffs, per-agent execution trees for subagents, per-turn token attribution across seven context categories, a project memory (MEMORY.md) viewer, and SSH support for inspecting remote sessions. The tool requires zero configuration, no API keys, and does not wrap or interfere with Claude Code. It runs on macOS, Linux, and Windows, and can also be deployed as a standalone Docker server with no outbound network calls.
claude-devtools parses JSONL session logs from ~/.claude/ to reconstruct full interaction traces. It resolves tool calls (Read, Edit, Bash, etc.) into their exact file paths, regex patterns, matched lines, and diffs, and visualizes token consumption per turn across seven categories: CLAUDE.md (global/project/directory), skills, @-mentioned files, tool I/O, thinking, team overhead, and user text. It shows compaction events when the context window is filled, renders subagent execution trees with tool traces, tokens, duration, and cost, and supports SSH remote sessions via ~/.ssh/config with agent forwarding. It also surfaces project memory layers from ~/.claude/projects/.../memory/, rendered with frontmatter and [[wikilinks]], and provides one-click copy for any message or code block plus full-session export to Markdown, JSON, or plain text. A Cmd+K command palette enables cross-session search.
- When Claude Code says 'Read 3 files' but you need to see exactly which files and their contents, use claude-devtools to inspect the actual file paths and highlighted lines.
- When your Claude Code session seems to forget things due to context compaction, use the compaction visualization to see what triggered the compression and what was lost.
- When debugging a multi-agent workflow or a subagent failure, expand the execution tree to trace each agent's tool calls, tokens, duration, and cost.
- When you need to share Claude Code output with a colleague without losing formatting, use the one-click copy or export to Markdown to produce clean, selectable text.
- When you manage servers remotely, use the SSH feature to inspect Claude Code sessions on any machine without logging into the server directly.
What are this agent's strengths and limitations?
- Zero-configuration: reads existing logs without API keys or modifications to Claude Code.
- Visualizes the context window with per-category token attribution, helping you identify what consumes tokens.
- Provides per-agent execution trees that reveal subagent activity beyond the final summary.
- Includes a project memory viewer with wikilinks and editor launchers, saving time navigating ~/.claude memory files.
- Tightly coupled to Claude Code's log format and features; no support for other AI assistants.
- Dependent on Claude Code v2.1.20+ where detailed tool output is hidden by default.
- Desktop app requires Electron; some users may prefer a lightweight CLI or web-only tool.
- Windows users may encounter SmartScreen warnings during installation.
How do you install or deploy this agent?
On macOS, install via Homebrew: brew install --cask claude-devtools, or download the .dmg from GitHub Releases (arm64 for Apple Silicon, x64 for Intel) and right-click → Open on first launch. Linux users can choose .AppImage, .deb, .rpm, or .pacman packages; Windows users run the .exe installer (click 'More info' → 'Run anyway' if SmartScreen appears). For Docker, run docker compose up and open http://localhost:3456, or manually: docker run -p 3456:3456 -v ~/.claude:/data/.claude:ro claude-devtools.
How do you use this agent?
After launching the app, it automatically scans ~/.claude/ for session logs—no configuration needed. Browse the list of sessions, click one to open detailed views of each message, tool call, token statistics, and context visualization. Use Cmd+K for cross-session search. The sidebar provides access to project memory; clicking a memory layer shows a rendered markdown view with frontmatter metadata, and you can open layers in your preferred editor via the 'Open in…' launcher. SSH remote sessions can be added via ~/.ssh/config. Each message and code block has a copy button, and you can export the session to Markdown, JSON, or plain text.