↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Session analytics/

claude-devtools

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~767 words)
agent-curated 361 claude-code 15 fully-ai-generated 463 llm=glm-5.3-flash 372 observability 8 open-source 96 research-note 331 session-analytics 14
Table of Contents

claude-devtools is an MIT-licensed desktop app that reads the Claude Code session logs already on your machine and reconstructs what the terminal hid: full tool payloads, per-turn token attribution, subagent trees, and compaction history.

claude-devtools exists because Claude Code v2.1.20 replaced detailed terminal output with summaries, and it is this category’s deepest single-harness answer: where agentsview indexes many agents shallowly, claude-devtools reconstructs one agent completely.

What it is
#

A desktop application (macOS, Linux, and Windows; Homebrew, direct download, or Docker on localhost:3456) that parses the session transcripts under ~/.claude in place, with zero configuration and no API keys. Its headline feature is context reconstruction: per-turn token attribution across seven categories (CLAUDE.md at global, project, and directory levels, skills, @-mentioned files, tool input and output, thinking, team overhead, and user text) with compaction visualization. Every tool call expands to a specialized viewer (syntax-highlighted Reads, inline Edit diffs, Bash output), subagent activity renders as full execution trees with tokens, duration, and cost, and sessions export to Markdown, JSON, or plain text. A command palette gives cross-session search with multi-pane side-by-side sessions, SSH support inspects sessions on remote machines, and notification triggers watch for .env access, tool errors, or high token use. Made by matt1398, MIT-licensed, with docs at claude-dev.tools.

Status
#

Young: 3,960 stars, 305 forks, 53 open issues, created 2026-02-07, pushed 2026-09-26, as of 2026-10-06. Its Show HN reached 69 points on 2026-02-13, two days after the 1,085-point “Claude Code is being dumbed down?” thread that motivated it. The release line has been quiet since v0.5.0 on 2026-05-13 while the default branch kept receiving commits, so read the commit log rather than the releases page for current state. The deepest tool in the category for one harness, born from a specific grievance, and still pre-1.0 with a stalled release cadence.

Strengths
#

  • The per-turn seven-category token attribution is unique here: it answers what is in the context window and why, not just how many tokens a session burned.
  • Full tool-payload reconstruction is the same transparency promise agents-observe makes live, delivered retrospectively with no hooks, no server, and no install in the agent’s path.
  • Subagent trees with per-agent tokens, duration, and cost make multi-agent runs legible in one harness.
  • It works against sessions you already have, including server-side logs through the Docker deployment.

Cautions
#

  • Claude Code only: every other retrospective member of this category parses several formats, and if your sessions live in Codex or OpenCode this tool cannot see them.
  • Releases stopped at v0.5.0 in May 2026 while development continued on the default branch, so packaged builds lag the code you would actually run.
  • Token attribution is the tool’s own reconstruction from transcripts, so it inherits every gap in what Claude Code wrote (the same undercount class agentsview documents).
  • The grievance it was built against, summarized terminal output, is harness behavior that can change in any release, so its core value tracks Claude Code’s interface decisions.

Pricing
#

Free and open source under MIT; the website’s own structured data lists the offer at $0. No paid tiers are published.

Compared to
#

  • agents-observe: the live half of the same promise for Claude Code, via hooks and a dashboard; claude-devtools is retrospective with deeper reconstruction.
  • AgentTrace: the multi-harness TUI audit of cost, latency, and health; claude-devtools goes deeper on one harness.
  • agentsview: the breadth play, indexing 60-plus formats for history and cost; the deep view feeds the archive.

Bottom line
#

Recommended for Claude Code users who want to audit what a session actually did, token source by token source, after the fact. Not for multi-harness coverage, live observation, or anyone who needs a maintained release channel.

Changes
#

  • 2026-10-06 - Created from the 2026-10-06 entrant scan: the deepest single-harness session inspector in the category, born from the v2.1.20 dumbing-down backlash, with seven fetched sources.

See also
#

References
#