AgentSight

See where your tokens go. Optimize your Claude Code workflow.

AgentSight is an open-source Rust CLI that parses your Claude Code session logs and surfaces usage analytics, cache efficiency metrics, and cost estimation — so you can work smarter, not more expensively.

Token attribution — Know which projects, sessions, and models consume the most tokens
Cache efficiency — Detect cache churning and context instability that waste tokens
Pattern detection — Find bash retry loops, exploration spirals, and subagent sprawl
Cost estimation — Optional dollar-cost breakdowns for pay-per-token API users

Install with Homebrew:

brew tap RealNerd/agentsight && brew install agentsight

Or with Cargo: cargo install agentsight

Quick Start

After installing, run agentsight diagnose in any project to get an instant efficiency report:

$ agentsight diagnose

 ── Diagnose: my-project ──────────────────────────────────
 Project:  work/my-project
 Tokens:   318,293 across 17 turns (35m)
 Cache:    70.6% hit ratio
 Cost:     $4.31

 ── Cache Stability ───────────────────────────────────────
 Classification: STABLE
 Cache front-loading looks healthy.

 ── Context Growth ────────────────────────────────────────
 Context size remained stable across the session.

 ── Tool Patterns ─────────────────────────────────────────
 No concerning tool patterns detected.

 ── Recommendations ───────────────────────────────────────
 Your session looks efficient. No changes needed.

Or let Claude Code do it for you. Install the slash command skill (not included automatically — you run this once after installing the CLI):

$ agentsight install-skill agentsight-diagnose

Then type /agentsight-diagnose inside Claude Code. It will run the analysis, read your CLAUDE.md, and suggest specific improvements.

Guides

Beginner

Install AgentSight, run your first diagnose, and build a weekly habit that keeps your Claude Code sessions efficient.

Intermediate

Understand every metric, optimize cache efficiency, compare sessions before and after CLAUDE.md changes.

Advanced

Cross-project benchmarking, model distribution analysis, the web dashboard, JSON scripting, and contributing.

What It Reads

AgentSight reads Claude Code's JSONL session logs from ~/.claude/projects/. It is read-only — it never modifies your sessions or configuration. It tracks:

Input, output, cache creation, and cache read tokens per turn
Tool usage (Read, Edit, Bash, Grep, Write, Glob, Task, etc.)
Session timestamps, git branch, and model used
Bash command retries and identical-error patterns

Open Source

AgentSight is dual-licensed under MIT and Apache-2.0. Contributions welcome.

github.com/RealNerd/agentsight