What I learned shipping four open-source Claude dev-tools in two weekends

---

title: What I learned shipping four open-source Claude dev-tools in two weekends

published: true

canonical_url: https://ferhatatagun.com/blog/four-tools-in-two-weekends

description: A meta post on the four-tool Claude dev-tool suite — why the SDK breaking was the constraint that made the work possible, one tool per insight, BYOK + browser-only.

tags: claude, anthropic, opensource, webdev

cover_image:

---

About a month ago I tried to import the Anthropic SDK into a Next.js project and the bundler crashed. The fix was straightforward — talk to the Messages API directly, ~150 lines of TypeScript replacing the SDK — but the side-effect was that I now had a hand-rolled SSE client lying around, with all of Claude's streaming behaviour visible to me at the protocol level for the first time.

That client became the seed of four small open-source tools, shipped over two weekends. Each one points a different microscope at the same protocol:

• [**claudoscope**](https://claudoscope-labs.vercel.app) — live x-ray of token economics: input, cache write, cache read, output, all visible as the response streams.
• [**agent-replay**](https://agentreplay.vercel.app) — paste a Claude agent trace, replay it step-by-step on a cinematic timeline.
• [**prompt-lab**](https://prompt-lab-promptly.vercel.app) — run two prompts (or models) on the same input, side by side, with output/cost/latency compared.
• [**tool-lab**](https://tool-lab-bice.vercel.app) — define Claude tools in a JSON editor, type the mock responses by hand, watch the agent loop play out.

All four run only in your browser, BYOK, no backend, MIT-licensed. Together they're around 400 KB gzipped; the shared SSE client is the same file in all four repos. Five long-form posts on [ferhatatagun.com/blog](https://ferhatatagun.com/blog) and Medium document the engineering decisions behind each one.

The work is done — the more interesting question for me now is what shipping them in this shape, on this timeline, taught me about building developer tools in the AI-tooling era.

**TL;DR**

• Resistance from the official SDK ended up being the most generative constraint. Without the crash, I would never have written the parser, and without the parser, I would never have noticed how much the SDK hides.
• "One tool per insight" beats "one tool for everything." Each of the four tools makes exactly one thing visible. They compose because they don't try to.
• BYOK + browser-only is a credibility multiplier. The threshold for "I'll try this" drops dramatically when there's no account to make and no server to trust.
• A `<150-line` shared protocol client across four projects is a more interesting reuse pattern than "extract into a library." It travels by copy-paste, but with intent.
• The articles are not promotion; they're scaffolding. Every tool needs a long-form artifact that explains *why* it exists, not what it does.

The constraint that made the work possible

If the Anthropic SDK had imported cleanly into my