Opening

Firebase's official Emulator Suite takes 30.8 seconds to start with data loaded; Firenook's own benchmarks clock in at 0.59 seconds. That gap earned it 167 points and 28 comments on Hacker News. Today we start with Firenook, then look at craft-anything, a Rust reverse-engineering engine that hit 17.7k★ in 30 days, plus a text-to-image model trained for roughly $21,000 with weights released straight under Apache-2.0. Whether that model is actually usable, I'll get to that in a bit.

Today's Open Source Picks

1. Firenook: a drop-in replacement for the Firebase Local Emulator Suite, no Java required

• What it is: If you work on Firebase backends, you know the routine: run firebase emulators:start, wait for Java to boot, then wait for a pile of services to warm up, and the moment you change something and restart, all your data is gone. Firenook is an Apache-2.0 licensed replacement that reuses your existing firebase.json, .firebaserc, SDKs, and the same ports (Firestore 8080, Auth 9099, Storage 9199, Functions 5001, Pub/Sub 8085, and so on). Basically you just swap firebase for firenook in your scripts. The services ship as native code inside a single npm package, no Java install needed, and it runs on Node 22／24／26.
• Recent buzz: 167 points and 28 comments on Hacker News. The site currently lists version 0.2.1, so it's still very early.
• Highlights: First, the project claims that under a two-hour stress test against a real app, median write latency was 5.1ms versus the official emulator's 208ms, with peak memory at 2.11GiB versus 22.57GiB. Second, add --keep-data and your Firestore／Auth／Storage data survives restarts, with a data-loaded startup of about 0.59 seconds (versus 30.8 seconds officially), and a crash costs you at most the last second of writes. Third, there's a built-in local console covering every service, running on the same port as the official Emulator UI, with Functions／Pub/Sub／Cloud Tasks／Eventarc pages that the official UI doesn't have. There's also firenook mcp, exposing 54 tools so agents in Claude Code, Cursor, or VS Code can query logs directly and ask why Security Rules rejected a given request.
• Getting started: npm install --save-dev firenook, run npx firenook init --adopt --dry-run first to confirm it recognizes your project, then swap firebase emulators:start for npx firenook emulators:start. For tests, use firenook emulators:exec "npm test".
• Versus alternatives: It goes head-to-head with Google's official Firebase Local Emulator Suite, and import/export uses the official format so you can switch back anytime. But the caveats up front: it does not replace firebase-tools, there's no deploy or login, and deployment still goes through the official CLI. Realtime Database, Hosting, App Hosting, and Data Connect are not emulated at all right now, so if your project leans heavily on RTDB, hold off.
• Who it's for: Frontend and backend engineers who fight Firestore／Auth／Functions local development every day and are done with Java and cold starts.
• Links: https://firenook.dev/ ／ https://github.com/firenook-dev/firenook

2. craft-anything: turning "study how they built it" into "tell an agent to rebuild it" as one pipeline

• What it is: A reverse-engineering engine written in pure Rust, structured as a four-step flow: Study → Build → Check → Ready. It first studies the original target with inspection tools, then has a coding agent rebuild it in the language and framework you specify, then compares the result against the original (websites get compared at three screen widths, games run self-tests) and iterates on fixes. It's aimed at tooling nerds, and also at anyone who wants to know how a particular interface was put together.
• Recent buzz: 17.7k ★ on GitHub, and it's a project less than 30 days old. Apache-2.0, primarily Rust.
• Highlights: The engine ships with 314 tools (298 of them in the engine core) covering native binaries.NET, Java, Android, Apple, Electron／JavaScript, websites, firmware, packet capture, databases, and game files. It bundles its own decoders, disassembler, and decompiler, so it runs without Ghidra installed; if you do have Ghidra, Deep mode will use it too. Every tool is both a CLI subcommand and an MCP tool, so you can craft-anything-cli mcp and feed it to any MCP client. Clean room mode is the default: it never reads source code or files, rebuilding purely from screenshots, visible text, and measured layout and color. Any tool that would peek inside is refused outright with a prompt to switch to Deep mode. Viewing the analysis itself doesn't require AI.
• Getting started: Requires Rust 1.97+. After git clone, run cargo run --release -p craft-anything for the desktop app, or cargo install --path apps/craft-anything-cli to install the CLI, then craft-anything-cli doctor to check your environment. If you'd rather not compile, GitHub Releases has signed installers for Windows／macOS／Linux. The quickest way to get a feel for it is craft-anything-cli rebuild <url> "Recreate this website".
• Versus alternatives: Unlike pure reverse-engineering tools like Ghidra or IDA, it doesn't just hand you pseudo-C and call graphs (Advanced mode has those too), it pipes the analysis into a coding agent to produce a runnable new project, then automatically scores similarity. But be clear-eyed: the README itself says early alpha, and honestly lists three buckets, "works now," "half done," and "coming next." On the build side it currently mainly supports a locally logged-in Claude Code; the Codex adapter hasn't been verified on real hardware, and account and usage come out of your own plan. It also explicitly states it's only for software you own or are permitted to inspect, and it won't bypass DRM, logins, or anti-cheat. Take that one seriously.
• Who it's for: AI application and tooling engineers who want reverse-engineering analysis wired directly into a coding agent and can live with alpha quality.
• Links: https://github.com/storytold/craft-anything

3. P-JiT (pjit-21k): a 2B text-to-image model trained by two people for about $21,000 in compute

• What it is: Field Notes published on November 4, 2026 by Linum's two co-CEOs, Sahil Chopra and Manu Chopra, releasing a 2B-parameter Pyramid-JiT (P-JiT) pixel-space text-to-image model with code and weights fully under Apache-2.0. If you want to self-host a generative model, or are just curious how far a small team can actually get, this is material you can pull down and run right now.
• Recent buzz: 267 points and 29 comments on Hacker News.
• Highlights: Architecturally, the authors say they shrank the attention context window by 4x and need 11x fewer samples to converge. The released model was trained on 660 million samples, and the authors put total compute cost at roughly $21,000, positioning it as a proof of concept on the way to larger video models. The site hosts 249 generated samples with full prompts, each clickable at original resolution, and the article itself includes a training configuration appendix and a Model Limitations section that doesn't dodge the weak spots. Credit where it's due.
• Getting started: Grab the Linum-AI/pjit-21k weights from Hugging Face and run them on your own GPU with the identically named model code on GitHub. If you don't want to set anything up, click through a few images in the site's gallery to see the prompts and actual outputs first, then decide whether it's worth your time.
• Versus alternatives: Compared to mature open-weight models like Stable Diffusion or FLUX, this is small and an experimental research release. The real selling point is that it fully documents a "$20k-tier" pixel-space training path, plus a commercially usable Apache 2.0 license, making it good for research and fine-tuning. Don't expect it to replace your current image-generation workhorse.
• Who it's for: AI engineers researching low-cost training paths, or anyone who needs a commercially licensed small model to fine-tune.
• Links: https://www.linum.ai/field-notes/pjit-21k ／ https://github.com/Linum-AI/pjit-21k ／ https://huggingface.co/Linum-AI/pjit-21k

Closing

Today's three picks share an obvious theme: they all challenge costs we take for granted in the big-vendor version, whether that's a Java runtime, the human effort of reverse engineering, or the budget for training an image model. If I had to recommend one, Firenook is the only thing here you could swap in at work tomorrow and roll back to the official tooling the moment it breaks. craft-anything and P-JiT are more like weekend toys: fun, but keep an exit plan. See you tomorrow.