8 Open Source AI Tools From GitHub Trending This Week

The biggest pattern on GitHub this week isn't a new model or another chatbot wrapper. It's the rise of "skills," modular capabilities you install into AI agents the way you install apps on a phone. Five of the top 20 trending repos across all languages are agent skills or skill infrastructure. But the week also delivered two tools that tackle the cost problem head-on, a vector index that fits 10 million documents in 4 GB, and a voice AI that runs entirely offline.

Here are eight repos worth your attention.

headroom: Cut your token bill by 60-95%

chopratejas/headroom — 27.6K stars, +10.6K this week

If you run AI agents in production, token costs are your most visible line item. Headroom compresses tool outputs, logs, RAG chunks, and conversation history before they reach the LLM. The claim is 60-95% fewer tokens with the same answers.

What makes it work isn't the compression itself. That idea has been around. It's the implementation. Headroom runs as a proxy server (zero code changes), a library, a CLI wrapper for Claude/Codex/Cursor, or an MCP server. It's reversible. headroom_retrieve to fetch the full uncompressed content when it needs it. And it mines failed sessions, writing corrections to your agent config files automatically.

If your monthly API bill is in the thousands, this is worth an afternoon of setup.

last30days-skill: Your agent's research department

mvanhorn/last30days-skill — 42K stars, +12K this week

The number one trending repo. It's an AI agent skill that searches Reddit, X, YouTube, Hacker News, Polymarket, and the web simultaneously, then synthesizes what it finds into a single brief. Results are ranked by actual engagement — upvotes, likes, and Polymarket odds — not by an algorithm guessing relevance.

Install it as a Claude Code plugin or via npx skills add and it works across 50+ agent hosts. For a sales team prepping for calls or a marketing team tracking competitor sentiment, this replaces hours of manual scrolling. The engagement-weighted ranking is the clever part: a Reddit thread with 3,000 upvotes surfaces above a blog post nobody read.

NVIDIA SkillSpector: Antivirus for AI agent skills

NVIDIA/SkillSpector — 5.3K stars, +3.7K this week

NVIDIA's research found that 26.1% of agent skills contain vulnerabilities and 5.2% show likely malicious intent. If you install community skills into your coding agent (and a lot of teams do), you're running untrusted code with access to your system.

SkillSpector scans skills before installation across 64 vulnerability patterns in 16 categories: prompt injection, data exfiltration, privilege escalation, supply chain attacks, system prompt leakage, and memory poisoning. It gives a 0-100 risk score with severity labels. Runs as a CLI or in Docker, no Python required.

This is the kind of tool that becomes mandatory the first time someone's agent gets compromised. NVIDIA shipping it gives it credibility that an indie project wouldn't have.

turbovec: 10 million documents in 4 GB of RAM

RyanCodrai/turbovec — 11.5K stars, +4.9K this week

Vector search is the backbone of RAG, and RAG is how most businesses put AI on top of their own data. Standard FAISS indexes eat memory. A 10-million-document corpus at float32 precision needs 31 GB of RAM.

Turbovec uses Google Research's TurboQuant algorithm to compress that same corpus to 4 GB, and searches faster than FAISS. Written in Rust with Python bindings, with hand-optimized SIMD kernels for ARM and x86. It ingests new vectors online with no retraining step. Runs entirely locally with no managed service, no data leaving your machine.

For anyone building RAG on a budget or in an air-gapped environment, this is the difference between needing a dedicated server and running on a laptop.

Agent-Reach: Give your agent the whole internet

Panniantong/Agent-Reach — 28.8K stars, +5.5K this week

Your AI agent is only as useful as the data it can reach. Agent-Reach gives any CLI-capable agent the ability to read and search Twitter, Reddit, YouTube, GitHub, LinkedIn, RSS feeds, and more. Zero API fees. Every backend has automatic failover, so when one scraping method gets blocked, it switches to another without intervention.

For lead enrichment, market research, or competitive monitoring, this removes the biggest friction point: configuring and paying for a dozen different platform APIs. The agent-reach doctor command tells you exactly what's working and what's broken.

graphify: Stop grepping, start querying

safishamsi/graphify — 67.2K stars, +5.5K this week

Graphify turns any folder of code, docs, PDFs, images, or videos into a queryable knowledge graph. Type /graphify in your AI coding assistant and it maps your entire project into a graph you can search semantically instead of grepping through files.

It works with 20+ AI coding tools: Claude Code, Codex, Cursor, Gemini CLI, Copilot, Aider. The output includes an interactive HTML visualization, a summary report, and a JSON file for programmatic access. It also runs as an MCP server for real-time querying.

The use case goes beyond code. If your organization has a document graveyard (years of wikis, specs, and reports scattered across folders), graphify turns that into something your AI agents can navigate.

CopilotKit: Build agent-powered UIs without starting from scratch

CopilotKit/CopilotKit — 35.1K stars, +1.5K this week

If you're building a product with AI agents and need a frontend, CopilotKit gives you the full stack: generative UI where agents render components dynamically, shared state between agent and UI, human-in-the-loop workflows, and multi-platform support (React, Angular, Vue, React Native, Slack, Teams).

The AG-UI Protocol they created has been adopted by Google, LangChain, AWS, and Microsoft. That kind of adoption pattern usually means the protocol is solving a real problem. In this case, standardizing how agents talk to frontends.

For teams building AI-integrated customer-facing apps, CopilotKit removes months of plumbing work.

Open-LLM-VTuber: Voice AI that never phones home

Open-LLM-VTuber/Open-LLM-VTuber — 11.3K stars, +1K this week

A fully offline voice AI with visual perception, Live2D avatars, and voice interruption. It supports local LLMs via Ollama, multiple ASR engines (Whisper, FunASR), and a range of TTS options. No internet required. MIT-licensed.

For healthcare, legal, or any regulated industry where sending voice data to a third-party API isn't an option, this is a starting point for building voice-based intake tools that keep everything on-premise. The desktop pet mode with transparent background works for customer-facing kiosk scenarios.

It's rough around the edges compared to commercial voice AI platforms. But if you're prototyping a voice agent and want to understand the architecture before committing to a vendor, start here.

What ties these together

Three of these eight repos (last30days-skill, Agent-Reach, and graphify) solve the same problem from different angles: getting real-world data into AI agents. Two others (headroom, turbovec) attack the cost of running those agents. NVIDIA's SkillSpector addresses the trust problem that comes with installing third-party code. And CopilotKit and Open-LLM-VTuber handle the interface layer.

The model layer is commoditizing. The interesting work has moved to data access, cost optimization, and security.