The Journal
·13 min read

Building Developer Tools That Actually Help

I built Superstack to stop paying for Claude's noise, and CodeHandoff to stop losing my AI sessions between Macs. A first-person case study on both — the problems, the architectures, the numbers, and what shipping a million small developer tools taught me.

Frequently asked

Frequently asked questions

What is Superstack?
Superstack is a local runtime layer that sits between your IDE/CLI and Claude. It compresses outbound context, caches repeat content, and routes simple prompts to cheaper models — producing 38–53% average output token reduction and 89–98% A/B win rates across Claude Haiku 4.5, Sonnet 4.5, Opus 4.6, and Opus 4.7, with quality measurably improved.
How much does Superstack cost?
Superstack is free forever for solo developers. Teams pay for shared configs, fleet analytics, and per-developer spend dashboards. The install is a single command — curl -sSL superstack.now/i | bash — and works with Claude Code, Claude Desktop, Cursor, Aider, Cline, Continue, Zed, RooCode, Goose, and OpenClaw.
How does CodeHandoff sync Claude sessions between Macs?
CodeHandoff uses an architecture I call Shadow Sync. Claude keeps writing to its native ~/.claude folder. A FSEvents watcher mirrors changes to a synced shadow copy in real time, and rclone keeps the shadow copies on every Mac eventually consistent. A small SQLite ledger resolves conflicts. When you sit at a second Mac, your sessions are already there — no rebase, no merge, no manual intervention.
Is CodeHandoff a subscription?
No. CodeHandoff is $9.99 one-time, unlimited Macs, lifetime updates. The purchase price is roughly the same as one month of lost productivity when your AI session is on the wrong machine.
What stack did you build Superstack and CodeHandoff in?
Superstack is a TypeScript daemon that runs locally on your machine, intercepting LLM traffic via a small proxy. CodeHandoff is a native macOS app — Swift for the UI/agent, Rust for the sync core, FSEvents for change detection, rclone for transport, and a LaunchAgent for background operation. Both ship as a single binary with no Electron or Docker overhead.
How do I verify the 38–53% token reduction claim?
The methodology and harness are documented at superstack.now/methodology. I A/B tested 47 prompts per model with and without Superstack, measured token counts and response quality via paired scoring, and re-run the suite on every Claude model release. Raw numbers and prompts are published.