Back to browse
Argus – VSCode debugger for Claude Code sessions

Argus – VSCode debugger for Claude Code sessions

by lydionfinance·Mar 3, 2026·2 points·1 comment

AI Analysis

●●SolidSolve My ProblemShip It

Debugger for Claude Code session costs—step-by-step breakdown of tokens and tool calls.

Strengths
  • Reads .claude directory without code changes; zero friction for Claude Code users.
  • Real-time session analysis with cost breakdowns identifies waste (40% duplicate reads, retry loops).
  • Rich visualizations (flow diagrams, cache hits, token tracking) make opaque AI behavior legible.
Weaknesses
  • Niche audience limited to Claude Code users; marginal if you don't use that specific IDE feature.
  • Only 4 installs at post time suggests adoption friction or discoverability problem.
Target Audience

Developers using Claude Code heavily who want to understand and optimize API spend.

Similar To

Langsmith · Loom (LLM debugging)

Post Description

I've been using Claude Code heavily and kept wondering "why did this session cost $2?" or "why did it read the same file 15 times?".

Argus is a VSCode extension that analyzes Claude Code sessions from the .claude directory and gives you:

- Step-by-step breakdown of what Claude did - Cost analysis (which tools burned the most tokens) - Performance insights (retry loops, duplicate reads, context pressure) - Token usage visualization (cache hits, compaction events) - Flow diagrams showing file dependencies

Extension: https://marketplace.visualstudio.com/items?itemName=argus-cl...

It's basically a "time machine debugger" for your Claude Code sessions. You can click through each step, see tool inputs/outputs, and understand exactly what went wrong (or right).

Tech stack: TypeScript, React 19, Chart.js, Vite

Some interesting findings from my own sessions: - 40% of cost was from reading the same files repeatedly - Retry loops cost ~$0.15 per loop on average - Cache hit ratio varies wildly (10% to 80%)

Limitations: - Only works with local .claude directories (not remote sessions) - Session format is undocumented, might break in future - Analysis rules are heuristic-based

GitHub: https://github.com/yessGlory17/argus

Would love feedback, especially from other Claude Code users!

Similar Projects

AI/ML●●●Banger

ContextSpectre – Reasoning hygiene layer for Claude Code sessions

Cuts 1.5M tokens from bloated Claude sessions with compaction archaeology and cost-per-decision tracking.

Solve My ProblemBig BrainDark Horse
ppiankov
102mo ago