Your app's network traffic, ready to paste into an LLM. Without burning your context window.

---

You're debugging with Claude, ChatGPT or Cursor, and the model asks the inevitable question: "what does the API actually return?"

Your options today are all bad. A HAR export weighs megabytes, full of headers, cookies and base64 noise. Copying responses by hand loses the flow: which click fired which call? And both happily leak your tokens and passwords.

NetDigest is one extra DevTools tab. It captures the traffic you already see, strips everything an LLM doesn't need, and hands you a digest that reads like a story: what the user did, what the app called, what came back.

How much smaller?

HAR export (full)         ████████████████████   100%   ~72,400 tokens
Raw JSON bodies           ███████████████░░░░░    75%   ~54,500 tokens
NetDigest (TOON, M)       ███░░░░░░░░░░░░░░░░░    15%   ~11,000 tokens
  + “1 per endpoint”      ██░░░░░░░░░░░░░░░░░░    12%   ~8,400 tokens

Measured on a realistic 32-request session: big lists, a 456-key settings dump, duplicate polling, query variants. Run pnpm bench to reproduce it. A spot check from a real app: a 3.4 MB /users response becomes ~60 lines that still show every key.

The structure survives the diet. Every object keeps all of its keys. What gets cut (long strings, repeated array items, deep nesting) is replaced by markers like …[+147 items, 150 total], and every export starts with a short preamble that teaches the model how to read them.

What it looks like

- type: marker
  label: "click button \"Sign in\""
- startedDateTime: "2026-06-12T01:21:28.430Z"
  method: POST
  url: "https://api.example.dev/api/login"
  status: 200
  initiator: "src/auth/LoginForm.tsx:41:20 (LoginForm.useMutation[signIn])"
  requestBody:
    email: jo@example.dev
    password: ***
  responseBody:
    success: true
    data:
      user: { id: 7, name: Jo Dev }
      token: ***

The click, the hook that fired the request, the payload with secrets already gone, the response. All in token-efficient TOON.

Features

• 🪶 Marker-aware compaction. Strings, arrays and depth get truncated; every key survives. Re-truncation always reports the original totals.
• 🔁 Dedup and diff. Identical calls merge into one entry. After the first array item, you only see the fields that changed.
• 🕵️ Always-on redaction. Passwords, tokens, API keys and JWT-looking values become ***. Auth headers and cookies are never captured in the first place.
• 🎬 Flow recorder. One click arms it: clicks, form submits and SPA navigations become timeline steps between the requests they triggered.
• 🧭 Real initiators. Source maps resolve every call back to src/hooks/useArticles.ts:18 (useArticles.useQuery) instead of a hashed chunk.
• 🗺️ API map. One click condenses the session into a normalized endpoint contract: methods, statuses, query keys, example shapes.
• 🎚️ Token budget control. S/M/L detail levels, a live token counter, per-endpoint overrides.
• 🧹 Curation. API-only, same-domain, flow-only and 1-per-endpoint filters, search, multi-select, Del to drop rows.

Quick start

pnpm install
pnpm wxt prepare   # pnpm 10 blocks postinstall scripts, run once
pnpm dev           # opens Chrome with the extension loaded

Or pnpm build and load .output/chrome-mv3 as an unpacked extension.

1. Open your app, hit F12, find the NetDigest tab. Capture runs from the moment DevTools opens.
2. Click Record, use your app, watch the story build itself.
3. Cherry-pick endpoints with Ctrl+click, pick a detail level, Copy TOON.
4. Paste into your LLM. That's it.

🖥️ Try it without installing: djdevpro.github.io/net-digest embeds the real, unmodified panel running on scripted demo traffic (locally: pnpm landing && npx serve landing/dist).

How the compaction thinks

Level	Strings	Array items	Depth	Use for
S	64 chars	1	3	API surface overviews
M (default)	255 chars	3	6	day-to-day debugging
L	1024 chars	8	10	deep dives

One endpoint deserves more? Select it, override items / depth / chars just for it, and the setting sticks for good.

Privacy

No backend, no analytics, no telemetry. Everything happens inside your DevTools session and only leaves your machine when you press copy. Details in SECURITY.md.

Contributing

PRs are welcome. The codebase is small and mapped out in CONTRIBUTING.md: dev setup, architecture, and the two invariants to respect. Release notes live in CHANGELOG.md.

License

MIT © Youcef Djidjelli