Opening

magpie hit 5.4k stars on GitHub in a single week. Released under the MIT license, it gathers all the scattered API keys from AI CLIs like Claude Code, Codex, and Gemini CLI into one menu bar utility. In this episode we'll also look at SCM, a fully offline macOS search tool that scored 173 points on Hacker News, and AIHOT, an open-source intelligence aggregator you can self-host with a single docker compose command. It's open source, sure, but we'll tell you what licensing landmine is hiding inside.

This Week's Open Source Picks

1. magpie: a local gateway that swallows every AI CLI's API keys and protocol differences

• What it is: Almost every engineer now juggles Claude Code, Codex, Gemini CLI, and a pile of API keys at the same time, and switching models usually means digging through several config files. magpie is a gateway that runs locally, letting you switch models and providers with a single click in the menu bar instead of hand-editing config files yourself.
• This week's buzz: 5,436 GitHub stars from a 30-day-old project, MIT licensed, built by yetone, the creator of avante.nvim, so it already carries a lot of community trust.
• Standout features: The gateway runs on 127.0.0.1:3425 and speaks four protocols at once, OpenAI Chat, OpenAI Responses, Anthropic Messages, and Gemini, translating between them, with streaming, tool calls, and reasoning all carried over correctly. Routing groups support strategies like smart (quota-based), rotate, and pace, so when one account's quota runs out it automatically switches to whichever account resets fastest. Existing Claude, ChatGPT, and Copilot subscription logins can be used directly as providers shared with other agents, no need to duplicate your keys.
• Quick start: curl -fsSL https://usemagpie.ai/install.sh | sh installs it as a menu bar app right away, or you can go install github.com/yetone/magpie@latest.
• Compared to similar tools: claude-code-router only serves the Claude Code client, and LiteLLM is a pure backend proxy that doesn't touch your local agents' configs. magpie's difference is that it does protocol translation and rewrites config files for 40-plus agents, while also letting you share subscription logins as providers.
• Who it's for: Anyone running two or three AI CLIs at once who's tired of juggling keys across config files.
• Link: github.com/yetone/magpie

2. SCM: runs fully offline, turning every photo and every video frame into searchable content

• What it is: "I'm sure I saved that screenshot somewhere but I can't find it," or "which minute of that demo mentioned that one line," SCM is built to solve exactly these pain points. It runs entirely on your machine with no API key required.
• This week's buzz: 173 points and 73 comments on Hacker News, 405 GitHub stars, MIT licensed.
• Standout features: Five search modes: Files does semantic search, Scenes segments videos and pinpoints the exact timecode, OCR supports 35-plus languages, Dialogue does precise phrase matching against Whisper transcripts, and Ask lets a local LLM answer questions against transcripts and OCR text with citations. The vision model runs on ONNX Runtime, defaulting to CLIP ViT-L/14@336. The README states it plainly: "your media never leaves the machine." No account required, no telemetry.
• Quick start: brew tap allenv0/scm, then brew trust allenv0/scm, then brew install --cask allenv0/scm/scm. The first launch downloads the model weights, and after that it runs completely offline.
• Compared to similar tools: Spotlight and Photos search rely on filenames and metadata, while SCM uses CLIP/SigLIP visual embeddings for semantic matching, and even segments videos into scenes so every frame is searchable. Compared to cloud photo album AI search, it requires no account and no uploading.
• Who it's for: Apple Silicon users who care about privacy and are tired of digging through mountains of screenshots and videos to find one thing.
• Link: github.com/allenv0/SCM

3. AIHOT: an open-source intelligence digest generator you can spin up with one docker compose command

• What it is: If you've ever wanted to build an industry intelligence site or a daily tech briefing for your team, you used to have to piece together RSS feeds, deduplication, translation, and ranking all by hand. AIHOT open-sources that entire pipeline under the MIT license, and you can get it running with a single docker compose command.
• This week's buzz: 6,170 GitHub stars from a 30-day-old project, written in TypeScript, MIT licensed.
• Standout features: A six-stage pipeline covering collection, pre-filtering, double independent scoring, writing, event clustering, and popularity ranking, with all prompts and thresholds exposed and editable. Popularity is scored per event rather than per article: the same source only counts once within 48 hours, and its weight is halved after 24 hours. It supports six source types, including RSS, web listings, JSON APIs, X/Twitter, and WeChat public accounts, defaulting to an 8 a.m. daily digest and a 10 a.m. Monday weekly digest.
• Quick start: git clone https://github.com/KKKKhazix/AIHOT.git myhot, then node scripts/init-env.ts --llm-key <你的 key>, then docker compose up -d --build. Open localhost:3000, with the admin panel at /admin.
• Compared to similar tools: Compared to a typical RSS reader, it adds AI double-scoring and event clustering. Compared to a newsletter generator, it tracks event relationships and popularity decay across days rather than just producing single articles. One thing to note: while the code is MIT licensed, "the AIHOT name and logo are not covered by the license," so if you self-host it, remember to swap in your own branding.
• Who it's for: Anyone who wants to self-host an industry intelligence site or team tech digest without building RSS deduplication logic from scratch.
• Link: github.com/KKKKhazix/AIHOT

Closing

This week's three tools line up neatly with three everyday pain points: managing a mess of API keys, the exhaustion of searching for files, and the time sink of building an intelligence site, and all three are free at their core under the MIT license, with a low barrier to getting started. If I had to pick just one to try first, I'd start with magpie, you'll feel the difference the moment it's installed. See you next week.