🇬🇧 GitHub Trending — July 20, 2026
Today's GitHub Trending reveals a development ecosystem where AI agent infrastructure is the central theme. Of the 7 featured repositories, at least 5 are tools that agents and language models use to operate — gateways, persistent memory, and code intelligence. The remainder covers learning and voice.
---
1️⃣ diegosouzapw/OmniRoute ⭐ 1,343 today (20,786 total)
AI gateway with 268+ providers — the hottest of the day.
OmniRoute is an open-source (MIT) gateway exposing a single endpoint for 268+ AI providers and 500+ models — Claude, GPT, Gemini, Kimi K3, GLM, DeepSeek, and 50+ free ones. Works with Claude Code, OpenAI Codex, Cursor, Cline, and Copilot. Includes quota-aware auto-fallback, RTK+Caveman token compression saving 15-95%, MCP/A2A support, multimodal, and Desktop/PWA versions. Built by 500+ contributors.
Why it gained traction: The fragmentation of the AI model market created a real problem: providers go down, quotas expire, different models have varying prices and quality. OmniRoute solves this with a single, transparent abstraction guaranteeing code never stops due to a provider outage. The "Never stop coding" promise resonated strongly.
---
2️⃣ msitarzewski/agency-agents ⭐ 949 today (134,402 total)
Complete AI agency — 134K stars and counting.
Agency-agents is a framework orchestrating specialized AI agents — from frontend wizards to Reddit community ninjas, from whimsy injectors to reality checkers. Each agent comes with personality, processes, and defined deliverables, forming a true AI agency.
Why it gained traction: Numbers speak for themselves — 134K total stars. The project capitalized on the moment every developer wants to build multi-agent systems but few want to deal with orchestration complexity from scratch. Agency-agents delivers a ready-made structure with clear roles.
---
3️⃣ tirth8205/code-review-graph ⭐ 663 today (22,404 total)
Local-first code intelligence graph for MCP and CLI.
Builds a persistent map of the codebase — a code intelligence graph — that AI coding tools (via MCP or CLI) query to read only what matters. The description mentions benchmarked context reductions on code reviews and large-repo workflows.
Why it gained traction: One of the most acute problems with coding agents today is excessive context consumption — the agent loads entire files when it only needs a fragment. Code-review-graph solves this with intelligent local-first indexing, reducing tokens and speeding up responses.
---
4️⃣ jamiepine/voicebox ⭐ 610 today (43,840 total)
Open-source AI voice studio.
Voicebox is described as "the open-source AI voice studio": voice cloning, dictation, creation — all local and user-controlled. With 43,840 stars, it's one of the most popular audio/ML projects on GitHub.
Why it gained traction: Demand for voice AI is growing with voice agents, assistants, and generated content. Voicebox offers a complete voice toolkit without depending on paid APIs or cloud services, appealing to both developers and content creators.
---
5️⃣ rohitg00/ai-engineering-from-scratch ⭐ 501 today (40,205 total)
Learn, build, and ship AI from zero.
An educational course/monorepo teaching AI engineering from fundamentals to deployment. With 40K stars, it's one of the most popular AI educational repositories.
Why it gained traction: The AI market is thirsty for skilled engineers, and this repo offers a hands-on learning path building everything from scratch without excessive abstractions. The promise "Learn it. Build it. Ship it for others" is straightforward.
---
6️⃣ kvcache-ai/ktransformers ⭐ 360 today (18,620 total)
Flexible framework for heterogeneous LLM inference.
ktransformers allows running inference and fine-tuning of LLMs leveraging heterogeneous hardware — CPUs, GPUs, NPUs — flexibly. Focused on KV cache optimizations, the main bottleneck in transformer inference.
Why it gained traction: With the proliferation of local models and the need to run them on varied hardware (not just NVIDIA), ktransformers tackles inference efficiency innovatively, allowing parts of the model to run on different devices.
---
7️⃣ topoteretes/cognee ⭐ 303 today
Persistent memory for AI agents based on knowledge graphs.
Cognee is an open-source memory platform for AI agents. It provides agents with persistent long-term memory across sessions using a self-hosted knowledge graph engine.
Why it gained traction: AI agents without memory are useless for continuous tasks — every conversation starts from scratch. Cognee solves exactly this: state persistence between sessions via knowledge graphs, enabling agents to learn and retain information.
---
📊 Trend Reading
The common thread uniting all these projects is a boom around AI agent infrastructure. Not new models or ML frameworks — but the support layers:
- Gateways (OmniRoute) — routing calls across dozens of providers - Persistent memory (Cognee) — agents that remember past interactions - Code intelligence (Code-review-graph) — coding agents that understand the codebase - Multi-agent orchestration (Agency-agents) — systems with multiple specialized agents - Efficient inference (ktransformers) — running models locally - Voice (Voicebox) — AI voice interfaces - Education (ai-engineering-from-scratch) — training the engineers building all of this
The pattern is clear: we're in the age of the agent platform. The tools enabling AI agents to operate autonomously — with memory, context, and resilience — are what the open-source market values most today.
Source: GitHub Trending
