finds.dev

// weekly digest · 5 picks

This week: reverse-engineered NotebookLM, hardened servers, and AI SOC triage

Five repos this week span a wide range of risk tolerance. One treats Google's undocumented internal APIs as a stable foundation (it isn't, and says so). Another is a decade-old PHP platform still mid-migration to a newer architecture, documenting its own mess in public. In between: infra hardening roles that actually check themselves against compliance baselines, a security platform wiring LLMs into case triage, and a dialogue-scripting compiler that's quietly shipped in several well-known indie games.

What ties them together is that none of them are toys. Each one asks something real of you before you get value out of it — reading the defaults before applying a hardening role, understanding a custom language before writing dialogue, accepting that a wrapper around private APIs can break on Google's schedule. Worth it, if you go in with eyes open.

  1. 01notebooklm-py
  2. 02ansible-collection-hardening
  3. 03agentic-soc-platform
  4. 04YarnSpinner
  5. 05PrestaShop

★ 20,073PythonMIT

An unofficial Python/CLI/MCP client for Google's NotebookLM (now Gemini Notebook) that reverse-engineers the web and mobile APIs to expose nearly every feature of the product programmatically, including things the web UI doesn't surface like bulk artifact export. It's aimed at developers and AI agent builders who want to drive NotebookLM as a grounded research/content-generation backend from scripts, CI, or Claude Code/Codex rather than clicking through the browser.

This is the most complete unofficial NotebookLM client you'll find, full stop. It covers notebooks, sources, chat, notes, labels, research agents, sharing, and every studio artifact type (audio, video, slides, quizzes, flashcards, reports, mind maps), and it exposes all of that consistently across a Python API, CLI, MCP server, and REST server. Two independent transport backends — the standard web protocol and a separately reverse-engineered Android gRPC path with checked-in protobufs — means there's a fallback when one breaks, which is the kind of redundancy you rarely see in scraping projects.

The auth tooling is also more thought-through than usual: Playwright login, browser-cookie import, and a master-token mode built for headless, unattended use with self-healing cookie refresh. There's a fault-injection harness, nightly qualification runs, and VCR-style cassette scrubbing in CI — real reliability engineering for something built on APIs Google never promised to keep stable.

What to know going in: the README itself calls this best for prototypes, not production, because it depends entirely on undocumented RPCs and a reverse-engineered mobile protocol that Google can change without notice. The master-token and cookie-import auth paths mean this library holds full-account credentials, not a scoped key — treat that credential store like you would a password vault. And the default install drags in Playwright plus a ~170MB Chromium download if you're not using cookie import or master token.

★ 5,524JinjaApache-2.0

A set of four Ansible roles (OS, SSH, nginx, MySQL) that apply hardening configurations mapped to DevSec's InSpec compliance baselines. It's for ops/infra people who want a config-as-code hardening pass across a mixed Linux fleet, rather than a security scanner or audit tool.

The thing that separates this from a pile of hardening defaults is that it ties back to verifiable baselines — linux-baseline, ssh-baseline, mysql-baseline, nginx-baseline — so you can run InSpec afterward and actually confirm the hardening took, instead of trusting the role's word for it. CI backs this up with real Molecule convergence tests on VMs across CentOS Stream, AlmaLinux, Rocky, Debian, Ubuntu, and even FreeBSD/OpenBSD for the SSH role, so the OS matrix in the README isn't aspirational.

argument_specs.yml on each role also means you get input validation and ansible-doc/IDE support, rather than reverse-engineering variable names from defaults/main.yml like a lot of role collections make you do.

What to know going in: only OS, SSH, nginx, and MySQL actually work — apache_hardening and windows_hardening are explicitly marked in-progress and not working, and there's nothing here for Postgres, Docker, or Kubernetes. The SSH and OS roles touch ciphers, kex algorithms, PAM stacks, and sysctl values that can break existing automation or lock you out of a box — read the defaults before you apply this, it is not a safe blind-apply. It also requires Ansible 2.16+, which rules out older pinned AWX/Tower setups, and the standalone per-service roles it replaced are now archived, so migrating off those means expecting some variable renaming.

★ 1,223Python

An open-source SOC platform that wires LLM-driven triage and investigation into a fairly complete SOAR/SIEM stack: alert ingestion, case correlation, IOC enrichment, playbooks, and a CLI that lets Claude Code/Codex talk directly to case data. Aimed at security teams who want AI-assisted alert triage without building the Django/Celery-style plumbing themselves.

The domain separation here is clean for a Django app of this size — alerts, artifacts, cases, enrichments, knowledge, and playbooks are each their own app, and the LLM-driven work (case analysis, module execution, playbook runs) is pushed into dedicated worker processes rather than blocking request handlers. The SIEM layer is a genuine abstraction over Splunk and ELK rather than a single-vendor integration with extra steps, and folding traditional SOAR actions and LLM analysis into the same Playbook system avoids running two disconnected automation engines side by side. Exposing case search, threat intel, and playbook authoring to coding agents via a CLI is a real, somewhat novel integration point rather than a chatbot stapled onto a dashboard.

What to know going in: every triage and investigation path runs through a per-alert LLM call, and there's no visible batching, caching, or backpressure for the alert-flood scenario the README itself pitches as the use case — that's exactly where per-alert LLM calls get expensive fast. Threat intel and CMDB integrations are pluggable via provider classes, but it's unclear how many real providers ship versus stub interfaces, so your actual integration cost is unknown until you try it. The project is still shipping schema and migration changes every few weeks, CodeQL/dependabot were only recently added, and there's no visible secrets-management or multi-tenant isolation story — worth weighing given this ingests raw security telemetry.

★ 2,870C#MIT

Yarn Spinner is a dialogue-scripting language and compiler for writing branching conversations in games — a screenplay-like syntax that writers can use directly, compiled down to bytecode a game engine runs. This repo is the engine-agnostic core (lexer/parser/compiler/runtime); most people will actually consume it through the separate YarnSpinner-Unity package.

This is a real compiler, not a templating trick — ANTLR-generated grammar, a constraint-based type solver with inference, and a numbered diagnostics system (YS0001 through YS0064) with documented examples for each error. It's also been proven in shipped commercial games (Night in the Woods, A Short Hike, Dredge, Venba), which is rare validation for a tool this size. The language keeps evolving too — node groups, smart variables, saliency strategies for picking which line to surface — so this isn't a frozen 1.0 that stopped growing after launch.

The engine-agnostic core is also cleanly separated from the Unity integration, so a Godot port or custom-engine host is plausible without forking the language itself.

What to know going in: in practice this repo is Unity-adjacent — other engines get no first-party integration, so you're writing your own VM host and asset pipeline against this API directly. It's maintained by a two-person commercial studio monetizing through Itch, Asset Store sales, and Patreon, which is fine today but ties project continuity to that revenue rather than a broad maintainer base. Learning Yarn syntax and concepts like node groups and smart variables is a real cost for writers on a small team, and there's no built-in localization workflow — wiring this into Crowdin or POEditor is entirely on you.

★ 9,243PHPNOASSERTION

PrestaShop is a full-featured open-source e-commerce platform written in PHP, used by merchants and agencies to run online stores with a Symfony-based back office. It's for people who want a self-hosted, extensible storefront rather than a SaaS like Shopify, and are prepared to run a real LMS stack (PHP, MySQL/MariaDB) to do it.

The module/hook/theme system here is a genuine extension architecture — you can override core behavior without forking, which is the whole point of picking a self-hosted platform over SaaS. The project has been mid-migration from legacy ObjectModel code to Symfony/CQRS for years and documents that transition in real detail in the repo, which is an unusual level of transparency for a codebase this old and this large. CI backs it up seriously: PHP unit and integration tests plus both Behat and Playwright UI suites run on every PR, and the Docker dev environment gets you a working install with seeded admin credentials in one command — a genuine convenience for a project this size.

What to know going in: you're working in two architectural paradigms at once, legacy ObjectModel/Smarty alongside newer Symfony/CQRS modules, and which one you hit depends on which part of the admin you're touching — the migration isn't finished, and that affects how predictable the codebase feels while you're in it. It's locked to PHP 8.1+ and MySQL-family databases, no flexibility there. As a widely-deployed e-commerce platform handling payment data it's a standing target for attackers — there's an active bug bounty because critical issues do surface, so staying patched isn't optional. And once a store is built out with marketplace themes and modules, migrating off PrestaShop later is a real project, not a config change.

If you want this kind of rundown in your inbox instead of hunting for it, the email signup on finds.dev takes one line and no spam you didn't ask for.

// past digests

Every public digest stays up, picks and takes as they were published.

The archive →