October was the month coding agents got cheaper, local models got bigger, and "agent skills" turned into a genre of their own. We went through every repo created last month that picked up real traction and kept the ten that AI developers should actually look at. Each one has a short note on what it is, who it's for, and our honest opinion.
Trend of the monthReverse Engineer Anything
The loudest story of the last few weeks wasn't a single new repo. It was a wave. REA, an MCP server and CLI that lets coding agents inspect binaries, Electron apps, APKs and firmware and then rebuild what they find, kept climbing all month. A flood of projects followed, each claiming that an agent took apart some closed app and rebuilt it as open source.
The best way to follow that wave is AIshihzong/awesome-reverse-engineer-anything (CC0, created Oct 13). It's a bilingual awesome list that tracks every notable case, groups the open-source rebuilds by the app they replace, and checks which projects actually used REA. Its finding: of the six most-shared cases, only one says in its own repo that it did. It also has a clear legal and controversy section, which matters a lot here.
AIshihzong / awesome-reverse-engineer-anything 608 PythonCC0-1.0Open
Bookmark the list, not the hype. Most "AI rebuilt X" posts don't hold up on a close read, and this is the one place that does that checking for you.
1. TokenEconomizercheaper coding agents with no agent changes
microsoft / TokenEconomizer 5.6k PythonApache-2.0Open
A research project from Microsoft Research and UC Berkeley. It learns from the sessions your coding agent has already run in a repo, then cuts waste from the sessions that follow. For example, it remembers that this repo needs python -m pytest instead of pytest, so the agent stops rediscovering it. The paper reports an 18–31% lower mean cost on SWE-Bench tasks across five agent/model combinations, and accuracy went up in four of the five.
If you pay for Claude Code or Codex by the token, this is the most practical repo on the list. It's a black box around the agent, so you don't have to switch tools to try it.
2. Harness²agents that improve their own harness
google-research / harness2 23.4k PythonApache-2.0Open
The "harness" is everything around the model: prompts, tools, memory, skills, orchestration. Harness² refines that harness while it works on a task. It proposes edits, probes their effect, and composes a better harness without ground-truth feedback. It also carries lessons forward to later tasks. The model weights never change.
This one is research, not a product. But the idea of tuning the harness instead of the model is where most agent gains are coming from right now, so it's worth reading even if you never run it.
3. DeepInferDeepSeek's production inference engine
deepseek-ai / DeepInfer 25.8k PythonApache-2.0Open
This is the framework behind DeepSeek's own large-scale serving, open-sourced for offline inference of DeepSeek-V4.1-Flash. It includes FlashMLA/DeepGEMM/DeepEP kernels, MoE expert parallelism, draft-token decoding, batch-invariant inference, deterministic sampling and structured outputs.
Most people won't run it, since it's built for clusters. If you serve MoE models at scale, though, reading how a frontier lab does it in production is worth a lot.
4. Project Maya300B+ MoE models on your own GPUs
mw00 / project-maya 2.8k C++MITOpen
Maya runs GLM-5.3-Flash (321B), DeepSeek-V4.1-Flash (748B) and MiniMax-M3.1-Flash (428B) on consumer hardware. It keeps your conversation's most-used experts on the GPU, the next ones in RAM and the rest on NVMe, and shuffles them as you chat. It exposes OpenAI- and Anthropic-compatible APIs. The README reports 30+ tokens/s on a single card and up to ~149 on a multi-GPU rig.
This is the counterpart to DeepInfer: datacenter models on a desktop. If you care about privacy or offline use, it's the most exciting local-LLM release of the month.
5. OXRa 1.3B document parser that turns PDFs into agent-ready data
SoMarkAI / OXR 3.8k PythonOpen
OXR turns PDFs and document images into Markdown or JSON with layout coordinates. At 1.3B parameters it claims #1 on olmOCR-Bench, OmniDocBench and DocParsingBench. Its layout module is pluggable, so you can swap in your own. There's a free online demo and API.
Document parsing is the boring, unglamorous step that breaks most RAG pipelines. A small, fast model that does it well is worth testing against whatever you use today. Read the license's extra terms before using it commercially.
6. Anthropic OSS Scannerfree vulnerability scanning for critical open source
anthropics / oss-scanner 1.8k DockerfileApache-2.0Open
Anthropic's enrollment repo for scanning important open-source projects for security vulnerabilities in isolated build environments. Maintainers get reproducible reports and proposed patches.
This isn't a tool you install. It's a service you sign your project up for. If you maintain something people depend on, it costs you nothing to enroll.
7. Floorplanan interactive plain-English map of your codebase
haydenschmitty / floorplan 3.4k JavaScriptMITOpen
A Claude Code / Codex plugin that reads your codebase and draws an interactive map of it. Every part, file and web request gets a plain-English name, and every connection between them is shown. It's aimed squarely at people who built an app with AI help and can't answer "what breaks if I change this?"
It's a natural companion to Archify, which is already on curepo. Archify is better for designing systems; Floorplan is better for understanding the one you already have. The README even includes a comparison.
8. unslopmake vibecoded UI stop looking AI-generated
heydamir / unslop 13.4k JavaScriptCC-BY-4.0Open
A skill for Claude Code, Codex and Antigravity. It first finds the "slop": purple gradients, glows, emoji labels, shouty copy, uneven spacing. Then it picks a design direction based on your product and fixes the screen pass by pass, like a designer's redline review. It works with Next.js, React, Vue, Svelte, Tailwind or plain HTML/CSS.
Every AI-built landing page looks the same, and users notice. This is the cheapest fix we've seen. Pair it with diagram-design if your docs need the same treatment.
9. OpenFilmyour coding agent becomes a video editor
openfilm / openfilm 5.6k TypeScriptMITOpen
Ask Claude Code or Codex for a film, and OpenFilm has it write the film as web pages: one film.html, with each shot a frame(t) function using three.js, canvas, SVG, CSS or GSAP. A local browser-based Studio then lets you move, restyle, trim and cut everything by hand.
It's the same idea as HyperFrames on curepo, video as HTML, but with a real editor attached. That matters, because no agent gets the timing right on the first try.
10. earshotears for coding agents
Leonxlnx / earshot 4.4k PythonMITOpen
Agents can now write the code that renders a demo or a game, but they can't hear the result. earshot gives them that sense through MCP. It describes what a sound file sounds like, scores how well it fits a description, finds the best moment in a long recording, and produces a second-by-second listening transcript of a whole film. It runs on CPU only.
This is a niche, but a clever one. If you're using OpenFilm or HyperFrames for product videos, earshot is the missing piece for sound.
Honorable mentions
Four more October launches that didn't make the top ten but are worth a look:
wheresryan22 / anatomy 26.4k JavaScriptA Claude skill that explains a technical idea by building an interactive isometric machine whose mechanism mirrors the concept.
ARahim3 / DigUp 6.4k SwiftA desktop app for on-device semantic search across photos, PDFs, documents, recordings and videos, built on EmbeddingGemma 2.
amosblomqvist / learnharness 8.1k TypeScriptAn agent harness that measures what a learner already understands, plans a path to their goal and teaches step by step with a persistent knowledge graph.
deepseek-ai / dsh-superagent-prototype 22.1k TypeScriptDeepSeek's research prototype of a proactive coding agent that lives in one GitHub repository and takes work requests from issues and chat.
Found something we missed? Every repo on curepo is hand-picked. Browse the full collection or check back for the November roundup.