GitHub This Week: Agent Tooling Ate the Trending Page

The GitHub trending page this week doesn't look like it used to. No new JavaScript frameworks. No database projects. It's almost entirely agent infrastructure: compression layers, memory engines, knowledge graphs, research tools. The plumbing that makes AI agents actually work is being built in the open, and it's moving faster than anything I've seen from the big platforms.

Ten repos jumped out this week for anyone building or buying AI automation. Here's what they do and why they matter.

Token compression became a product

chopratejas/headroom picked up 14,272 stars this week, which is a lot for a library that most people had never heard of seven days ago. It compresses everything your AI agent reads (tool outputs, logs, RAG chunks, files) before it hits the LLM. Same answers, 60 to 95% fewer tokens.

It works as a Python/TypeScript library, a proxy you can stick in front of any agent with zero code changes, an MCP server, or a one-command wrapper for Claude Code, Codex, Cursor, Aider, and Copilot. The benchmarks show 47 to 92% token savings with zero accuracy loss on GSM8K and TruthfulQA. There's also a "headroom learn" feature that mines your failed agent sessions and writes corrections to config files. Boring sounding, until your agent bill drops by half.

Apache-2.0 license. 16,935 total stars.

Two code knowledge graphs, same week

colbymchenry/codegraph gained 8,661 stars this week (43,768 total). It builds a pre-indexed code knowledge graph that AI coding agents can query instead of running expensive grep-and-read loops. One tool call returns entry points, related symbols, and code snippets. Benchmarked across seven repos: 16% cheaper, 47% fewer tokens, 58% fewer tool calls. Supports 20 languages, recognizes routing patterns in 14 frameworks, runs entirely local on SQLite.

Lum1104/Understand-Anything gained 7,466 stars (54,377 total). Different approach: a six-agent pipeline scans your codebase, builds a knowledge graph of every file, function, class, and dependency, then gives you an interactive dashboard to explore it visually. The tree-sitter-plus-LLM hybrid creates reproducible structural graphs. It also generates guided architecture tours ordered by dependency, which is genuinely useful for onboarding. Works with Claude Code, Cursor, Copilot, Codex, Gemini CLI, and others.

Both MIT licensed. Two code knowledge graph repos trending in the same week says something: agents need better context, and grep is running out of runway.

Microsoft's document converter keeps growing

microsoft/markitdown added 13,359 stars this week, pushing to 147,347 total. It converts PDF, DOCX, PPTX, XLSX, images, audio, HTML, and basically anything else into clean Markdown. Plugin system for extending file types. Azure Document Intelligence integration for advanced OCR.

Quiet utility repo that ends up in every AI pipeline eventually. If you're feeding documents into an LLM for intake processing, RAG, or content extraction, you need something like this. MIT license.

Agents got memory and internet access

supermemoryai/supermemory gained 2,924 stars this week (26,015 total). It's a memory engine that extracts facts from conversations, tracks changes over time, resolves contradictions, auto-forgets expired info, and delivers the right context at query time. Ranked first on LongMemEval, LoCoMo, and ConvoMem benchmarks. Connectors for Google Drive, Gmail, Notion, OneDrive, and GitHub. Ships as an MCP server with plugins for Claude Code, Cursor, and most major coding agents.

If you're building an agent that processes intake or talks to customers, you want it to remember what it learned last time. Supermemory handles that without you building a custom memory layer.

Panniantong/Agent-Reach gained 2,289 stars (23,159 total). One-sentence install gives any AI coding agent access to 15 platforms: Twitter, Reddit, YouTube, GitHub, Bilibili, and others. Zero API fees. Cookie-based auth stored locally. Pluggable channel architecture.

If you've tried to get a coding agent to research something on Reddit or Twitter, you know the friction. Agent-Reach solves it with a single install command. MIT license.

AI research on autopilot

mvanhorn/last30days-skill gained 2,718 stars this week (31,012 total). It's an agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web, then synthesizes a summary scored by real engagement (upvotes, likes, prediction market odds). It resolves entities before searching (people to handles, companies to subreddits), merges the same story across platforms into one item, and outputs shareable HTML briefs.

The "Best Takes" section surfaces the funniest or most viral quotes. The comparison mode handles "X vs Y vs Z" queries automatically. There's an ELI5 mode for plain-language rewrites. Works on Claude Code, Codex, Cursor, and 50+ agent skills hosts. MIT license, 1,012 tests.

This one caught my eye because the scoring-by-engagement approach is different from standard search. It surfaces what people cared about, not what an algorithm wants you to see.

Voice and video tools

OpenBMB/VoxCPM gained 4,298 stars this week (27,509 total). VoxCPM2 is a tokenizer-free text-to-speech system with a 2B parameter model trained on 2M+ hours of multilingual speech. Supports 30 languages, outputs 48kHz studio-quality audio, and can create voices from text descriptions without reference audio. Fine-tune with as little as five minutes of audio. Production-ready via vLLM-Omni with an OpenAI-compatible endpoint. Apache-2.0 license.

The voice design feature (describe a voice in natural language, get that voice) is a genuine improvement over cloning-only approaches for voice AI builders.

harry0703/MoneyPrinterTurbo gained 7,992 stars this week (81,196 total). One-click AI short video generation: write a prompt, get a finished video with voiceover, subtitles, background music, and stock footage. Web UI and API. Portrait and landscape formats, multiple LLM providers, multiple TTS engines. MIT license.

81,000 stars tells you something about the demand. Small businesses want video content. They do not want to learn video editing.

The frontend layer for agents

CopilotKit/CopilotKit gained 1,753 stars this week (33,666 total). Frontend stack for building agent-native apps: generative UI, shared state between UI and agents, human-in-the-loop workflows. React and Angular. They created the AG-UI Protocol, now adopted by Google, LangChain, AWS, and Microsoft. The useAgent hook connects UI components directly to agent workflows. MIT license.

If you're building an intake form that an AI agent drives, this handles the UI layer so you don't wire up agent-to-frontend communication yourself.

The week in one sentence

The open-source ecosystem is building agent infrastructure faster than the platforms. Token compression, memory, internet access, code understanding, research, voice, video, and agent UI all shipped meaningful updates this week, mostly from individual developers or small teams. The tools exist. The question is which ones to wire together.