Agent Productivity-to-Cost Correlator — Is your AI coding agent actually shipping code?

Codelens AI ties Claude Code token usage to actual git output. It reads your local Claude Code session files, correlates them with git commits by timestamp, and serves a dashboard answering: "Am I getting ROI from my AI coding agent?"

• One command, zero config
• All data stays local
• Works with any git repo where you've used Claude Code

npx claude-roi

# npm
npm install -g claude-roi

# pnpm
pnpm add -g claude-roi

# yarn
yarn global add claude-roi

Then run anywhere:

claude-roi

git clone https://github.com/Akshat2634/Codelens-AI.git
cd Codelens-AI

# Install dependencies (pick one)
npm install
# or
pnpm install
# or
yarn install

# Run it
node src/index.js

• Node.js >= 18 — Download
• Git — installed and configured with user.name and user.email
• Claude Code — you must have used Claude Code at least once so session data exists at ~/.claude/projects/

npx claude-roi

This parses your ~/.claude/projects/ session data, analyzes your git repos, and opens a dashboard at http://localhost:3457.

claude-roi                        # default: last 30 days, port 3457
claude-roi --days 90              # look back 90 days
claude-roi --port 8080            # custom port
claude-roi --no-open              # don't auto-open browser
claude-roi --json                 # dump all metrics as JSON to stdout
claude-roi --project techops      # filter to a specific project
claude-roi --refresh              # force full re-parse (ignore cache)

The dashboard includes:

• Hero stats — total cost, commits shipped, cost per commit, ROI grade
• Smart insights — auto-generated observations about your usage patterns
• Cost vs Output timeline — dual-axis chart of daily cost and lines added
• Model comparison — cost breakdown by Claude model
• Session length analysis — which session sizes have the best ROI
• Productivity heatmap — GitHub-style grid showing when you're most productive
• Sessions table — sortable, expandable table with per-session metrics and matched commits

1. Parses JSONL session files from ~/.claude/projects/
2. Analyzes git history from each repo you've worked in with Claude
3. Correlates sessions to commits by timestamp (during session + 30min buffer)
4. Calculates cost using Anthropic token pricing (input, output, cache read/write)
5. Serves an interactive dashboard on localhost

Parsed session data is cached at ~/.cache/claude-roi/parsed-sessions.json. On subsequent runs, only new or modified JSONL files are re-parsed, making startup near-instant. Use --refresh to force a full re-parse.

Token costs are version-aware and calculated per model (see Anthropic pricing):

Line survival uses an approximate heuristic: if lines added in commit A are deleted by a subsequent commit on the same file within 24 hours, they're counted as "churned." This is not git-blame-based tracking and survival rates are rounded to the nearest 5%.

Codelens-AI/
├── package.json
├── README.md
├── .gitignore
└── src/
    ├── index.js          # CLI entry point
    ├── claude-parser.js  # Parse Claude Code JSONL session files
    ├── cache.js          # Parsed data caching layer
    ├── git-analyzer.js   # Parse git log with branch awareness
    ├── correlator.js     # Match sessions to commits by timestamp
    ├── metrics.js        # Calculate ROI metrics and insights
    ├── server.js         # Express server + API routes
    └── dashboard.html    # Single-file dashboard (inline CSS/JS)

Releases are automated via GitHub Actions. To publish a new version:

npm version patch   # or minor / major
git push --follow-tags

This automatically publishes to npm and creates a GitHub Release with auto-generated notes.

Setup (one-time): Configure trusted publishing on npm for the claude-roi package, linking it to the GitHub Actions workflow. No tokens or secrets needed.

Contributions are welcome! Please see CONTRIBUTING.md for development setup, guidelines, and ideas for contributions.

All data stays on your machine. The only network requests are for Chart.js and Inter font from CDNs. No telemetry, no data collection.

MIT