Opening

The hottest thing on Hacker News this week is fugleramme, a Raspberry Pi that listens to birdsong through a microphone and automatically renders it as an 1800s-style natural history illustration, racking up 2,318 points and 255 comments. Meanwhile, codex-with-chatgpt uses the ChatGPT subscription you're already paying for to handle planning and review, hitting 5.4k stars in 30 days. There's also m3e-canvas, which turns UI sketches into AI prompts. Let's dig into all of them.

This Week's Open Source Picks

1. codex-with-chatgpt: Let the ChatGPT subscription you're already paying for handle Codex's planning and review

• What it is: The most expensive part of AI coding usually isn't the file editing itself, it's the planning and code review conversation, which often burns through your API quota in a few back-and-forth exchanges. This tool routes that conversation back to your existing ChatGPT Plus/Pro subscription, leaving Codex to just execute the plan and edit files.
• This week's buzz: 5,448 stars on GitHub, hitting 5.4k as a new project within 30 days.
• Standout features: The ChatGPT side only pulls from nine read-only MCP tools (workspace_info, read_file, git_diff, test_status, etc.), reading only the specific lines it needs instead of uploading the whole repo. The two sides exchange minimal messages through an INIT→PLAN→EXECUTED→REVIEW→DONE state machine, and the review stage actually inspects the git diff and test logs rather than taking the agent's word that tests passed. It uses OAuth 2.1 with PKCE, dynamic client registration, and rotating refresh tokens; .env* and credential files are blocked by default, no API key is required, and the repo ships with 150 tests.
• Getting started: The laziest route is to paste the entire README install instructions to your Codex agent and let it install everything itself (cloudflared gets auto-detected). The manual route is running pnpm install && pnpm build to produce the c2c executable, running c2c setup, then managing it with c2c status / doctor / pair / logs.
• Compared to alternatives: Versus the native Codex workflow, the difference is that planning and review no longer consume API quota, instead using the subscription-based ChatGPT web app. Versus assistants that require uploading your whole repo or pasting it into a chat box, this one fetches lines on demand, keeping your code on your local machine at all times.
• Who it's for: Engineers who already pay for ChatGPT Plus/Pro and don't want to burn additional API quota on planning and review.
• Link: github.com/XiaoDuoYa/codex-with-chatgpt

2. m3e-canvas: Draw a Material 3 screen and get a ready-to-use AI prompt out of it

• What it is: The hardest part of vibe coding usually isn't writing code, it's "I can't quite describe what I want it to look like." This tool lets you sketch Material 3 Expressive screens directly in your browser, then automatically converts them into a well-structured natural language prompt you can feed straight to an AI to generate.
• This week's buzz: 7,420 stars on GitHub, hitting 7.4k as a new project within 30 days.
• Standout features: Built-in components drawn to the Material 3 Expressive spec, including buttons, FABs, chips, app bars, navigation components, cards, lists, dialogs, and text fields. You can link screens together, set "tap to navigate" or "swipe to navigate," and choose transitions (slide, fade, expand), then click through and preview the entire flow. Any given screen can be toggled between 412×892 mobile and 1280×800 desktop layouts, and this gets written into the prompt as "one screen, two widths." The generated prompt supports Chinese, English, Japanese, and Korean, and all data is stored in localStorage with no backend involved.
• Getting started: Just open lnkiai.github.io/m3e-canvas and start drawing, no signup or install required. To self-host, clone the repo and run npm install && npm run dev (localhost:3000), then npm run build to output static files to ./out.
• Compared to alternatives: The README itself points to matraic/m3e and Beer CSS, which give you Material 3 components or CSS but still require you to build the UI yourself. m3e-canvas flips this: instead of handing you code, it hands you a prompt that already includes the layout, navigation flow, and theme settings for the AI to generate from.
• Who it's for: Anyone doing vibe coding who struggles to describe what the UI should look like, and would rather communicate through a visual sketch before handing it off to AI.
• Link: github.com/lnkiai/m3e-canvas

3. fugleramme: An e-paper photo frame that listens for birdsong and renders it as an 1800s natural history illustration

• What it is: A Raspberry Pi paired with a microphone listens for birdsong outside your window, identifies the species, then lays them out as a classical natural history plate rendered on an e-paper display, all running fully on-device with no cloud involvement.
• This week's buzz: 2,318 points and 255 comments on Hacker News, the top-buzzed project this week; 3,050 stars on GitHub.
• Standout features: Identification is handled by BirdNET-Go, running in a separate container; Fugleramme periodically hits its API for results and focuses purely on presentation. The image library contains over 800 cutout images covering more than 400 bird species, all sourced from genuine classical natural history plates and hand-picked one by one; the README explicitly states "none of these were AI-generated." It only redraws when the detected bird species changes, conserving e-paper refresh cycles. Everything runs entirely on-device with no dependency on cloud services.
• Getting started: If you just want to see it in action without buying hardware, spin up docker run -d -p 8080:8080 -v fugleramme:/data -e FUGLERAMME_DETECTOR_URL=http://birdnet.local:8080 ghcr.io/arnegiacomo/fugleramme for the web version. For local development, run uv sync and open uv run fugleramme-fake-detector paired with uv run fugleramme-dev, which runs it with a mock detector. To actually build a physical frame, run install.sh on a Raspberry Pi 5 with a 13.3-inch Inky Impression e-paper display, a microphone, and an A4 photo frame.
• Compared to alternatives: Versus bird audio identification projects like BirdNET-Pi and BirdNET-Go, this isn't trying to replace the identification engine, it's a presentation layer built on top of one. Those give you a species list and spectrograms; Fugleramme gives you a classical illustration that changes its content based on whoever's visiting your window. Worth noting: the bird species data skews toward Northern Europe, the UK, and Central Europe, so accuracy for Taiwanese birds will be lower. The code is MIT-licensed, but BirdNET-Go's model is CC BY-NC-SA 4.0, non-commercial use only.
• Who it's for: Bird watchers, anyone who wants a home device that changes its own content, or engineers curious about a real-world example of "fully local, zero AI-generated imagery."
• Link: github.com/arnegiacomo/fugleramme

Closing

These three tools are really all about the same thing: cutting costs. codex-with-chatgpt cuts your API spend, m3e-canvas cuts the cost of "I can't describe what I want," and fugleramme cuts cloud dependency. If you can only try one, I'd start with codex-with-chatgpt, it'll hit home hardest for engineers burning tokens every day. New picks next week, see you then.