I’ve been there—hammering away on a feature, building something cool with Claude Code, and then bam—API quota exhausted. You’re either staring at a paywall or waiting for the next billing cycle. It’s the one headache that unites everyone using AI coding tools.
This week, a handful of open-source projects landed on GitHub that feel less like hype and more like honest fixes for real problems. Some of them are weirdly specific. Others are just clever. Let me walk you through the ones that caught my eye—especially the first one, because it might change how you think about AI coding costs.
The first project is 9router, and it’s been climbing the GitHub trending chart fast—gained nearly 2,000 stars in a week. The idea is straightforward: you plug it behind your AI coding tools (Claude Code, Cursor, Copilot), and it routes your API calls across 40+ providers and 100+ models. So you’re using Claude Code’s interface, but behind the scenes, it might be sending requests to free Gemini tiers or cheaper DeepSeek models.
It also has a three-layer automatic failover mechanism. When your paid subscription quota runs out, it quietly switches to a cheaper model. When that’s exhausted, it falls back to a free tier. You don’t get interrupted. There’s also a built-in RTK Token Saver that compresses outputs from tools like git diff and grep, cutting 20–40% of token usage per request. A “Caveman” mode can slash output tokens by another 65%.
Install it with npm install -g 9router, fire up a local dashboard at http://localhost:20128, and configure. Supports Docker, VPS, and Cloudflare Workers. If you’re burning through API credits every week, this is worth a look. (Repo: https://github.com/decolua/9router)
Next up is jcode—a Rust-based coding agent harness that’s all about performance. The numbers are eye-popping: memory usage is only 1/14th of Claude Code’s, and first render takes 14 milliseconds versus Claude Code’s 3.4 seconds.
The reason? It’s built from scratch—terminal rendering engine, Mermaid chart rendering (claimed to be 1,800× faster than the official mermaid-cli), everything custom. It also supports a Swarm multi-agent collaboration mode: you start multiple agents in the same repo, and they automatically coordinate, detect conflicts, and communicate. One agent writes frontend, another writes backend, no stepping on each other’s toes.
There’s a semantic memory system that does vector embedding after each conversation turn, retrieving relevant memories via cosine similarity in the next turn. The swarm mode is still experimental, but for devs obsessed with latency, this is a playground. (Repo: https://github.com/1jehuang/jcode)
From the performance extreme to the educational side, dive-into-llms comes from Shanghai Jiao Tong University. It’s a series of hands-on LLM programming tutorials, currently sitting at over 36,000 stars.
Covering 11 chapters—prompt learning and chain-of-thought, knowledge editing, mathematical reasoning, model watermarking, jailbreak attacks, LLM steganography, multimodal, GUI agents, RLHF alignment—each chapter provides slides, documentation, and runnable Jupyter notebooks.
It’s not academic fluff. You actually code. Recently, they partnered with Huawei’s Ascend to release a domestic GPU-based course, covering the full LLM development pipeline with PPTs, lab manuals, and video tutorials. No installation needed—just clone and start learning. (Repo: https://github.com/Lordog/dive-into-llms)
Let’s talk memory. agentmemory is a persistent memory system built specifically for AI coding agents—Claude Code, Cursor, Gemini CLI, Codex. Think of it as giving your agent a brain that remembers what you’ve said across sessions: project architecture, coding preferences, past decisions.
The memory architecture mimics human cognition with four layers: working memory, episodic memory, semantic memory, and procedural memory. Details you mention gradually solidify from temporary to long-term storage. For search, it uses a hybrid approach—BM25 + vector search + knowledge graph—with a claimed 95.2% recall rate on the top 5 results.
One-line setup: npx @agentmemory/agentmemory starts a memory server. If you’re tired of repeating yourself to your coding agent, this is your fix. (Repo: https://github.com/rohitg00/agentmemory)
3D Gaussian Splatting has been one of the hottest areas in computer vision for the past couple of years. But editing the resulting data has been a pain—until SuperSplat. Made by the PlayCanvas team, it’s a browser-based 3DGS editor. No downloads, no installs.
You can view, inspect, crop, merge, and optimize Gaussian Splat data. The rendering runs on WebGL/WebGPU, and the editing is surprisingly smooth. It also includes a video rendering feature to publish a flythrough of your scene directly to the web.
Just open supersplat.at/editor and dive in. Zero friction. (Repo: https://github.com/playcanvas/supersplat)
Ever wanted to turn your notes or technical articles into hand-drawn style illustrations? ian-handdrawn-ppt is a skill that does exactly that—generates pencil-sketch style diagrams with a clean, white background, thin lines, and soft blue/green highlights.
It doesn’t just slap drawings together. The system first understands your material, extracts the narrative structure, then maps it to a suitable layout—cover metaphor, left-right comparison, flowchart, matrix. You get a 21:9 cover image and 16:9 content images in PNG format, ready for slides or blog posts.
Perfect for making technical explanations more approachable, especially if you’re building course materials or documentation. (Repo: https://github.com/helloianneo/ian-handdrawn-ppt)
Finally, Pixelle-Video from Alibaba’s AIDC-AI team is a fully automated short-video engine. You give it a topic, and it handles everything—script writing, image generation, voiceover, background music, and final video composition.
Built on ComfyUI, every step is modular, so you can swap out the image generation model or TTS engine. Beyond basic video creation, it also supports digital human lip-sync, image-to-video, and motion transfer.
It’s still early, but the potential for content creators is obvious. (Repo: https://github.com/AIDC-AI/Pixelle-Video)
I’ve been following GitHub trending for years, and what I notice this week is a shift toward utility over hype. Projects like 9router and agentmemory solve friction points that every AI-assisted developer faces. jcode pushes the boundary of what a coding agent can do performance-wise. And the educational content from dive-into-llms shows that open source can be a powerful teaching tool.