Top 10 GitHub Repos: AI Videos, Gorgeous Diagrams, Token Savings and More

Originally published at https://tekmag.thsite.top/top-10-github-repos-ai-videos-gorgeous-diagrams-token-savings-and-more/

GitHub trending repositories reveal where developers are investing their time—agent memory systems, editorial diagram templates, local-first infrastructure, and token-cost routing lead the pack in August 2026.

Andrew Warner and Adam broke down the latest wave of open-source tools on their August 21 episode covering 57,449-view content about GitHub's top repositories. Eleven of the thirteen repos they discussed had never appeared on TekMag before. This piece covers the ten most notable among them.

Key Takeaways

• public-apis/public-apis leads at 474,008 stars, the longest-running curated directory of free APIs on GitHub.
• volcengine/OpenViking (34,849 stars) unifies agent memory, RAG knowledge bases, and skill management in one system.
• cathrynlavery/diagram-design (29,190 stars) offers 38 editorial diagram types designed specifically for AI coding assistants.
• modular/modular (29,421 stars) ships the Mojo programming language, which claims performance beyond Python while keeping Python compatibility.
• NVIDIA-NeMo/Switchyard (2,636 stars) routes LLM requests across models and providers to balance quality against token cost.

GitHub is a code hosting platform, and this month's trending repos show where developers are investing their attention.

OpenViking gives AI agents a persistent memory layer

volcengine/OpenViking is a self-evolving context database that unifies agent memory, knowledge retrieval, and skill management for AI applications. According to its repository description, the project sits at the intersection of Agent Memory, Knowledge RAG (Retrieval-Augmented Generation), and Skills orchestration. Released by ByteDance's Volcengine team, it attracted 34,849 stars as of early September 2026.

Agent memory remains one of the hardest problems in production AI. Most chatbots lose context between sessions, and RAG systems alone do not adapt to new information without manual re-indexing. OpenViking addresses both gaps by maintaining a living knowledge graph that evolves as the agent interacts with users and documents. The system stores conversational history, extracted facts, and procedural skills in a single queryable store.

The project is written in Python and includes a REST API for integration with existing agent frameworks. According to the GitHub repository page, the default configuration supports vector search, graph traversal, and structured metadata filtering out of the box. For teams already running Claude or GPT-based agents, OpenViking reduces the engineering effort required to add persistent memory from weeks to days, following the same pattern as MCP Server integration into developer tooling through standardized protocol layers.

Volcengine, ByteDance's cloud division, has been shipping AI infrastructure tools since 2023. This repo follows the same pattern as their earlier inference optimization projects: well-documented, production-ready, and aimed squarely at enterprises building agent workflows rather than hobbyists experimenting with prompts.

Diagram Design brings editorial-quality visuals to AI coding tools

cathrynlavery/diagram-design is a collection of 38 hand-crafted diagram templates that generate clean SVG output for Claude Code, OpenAI Codex, and similar AI pair programmers. The project hit 29,190 stars because it solves a specific pain point: AI-generated diagrams from Mermaid and PlantUML often look generic and visually inconsistent.

The templates use pure HTML and SVG with no shadow effects, no gradient clutter, and a consistent typographic scale. Each diagram type follows editorial design conventions used in publications like The Verge and Apple's developer documentation. The repo ships a single self-contained HTML file that AI coding agents can reference when generating architectural diagrams, sequence flows, or system architecture illustrations.

According to the project README, the author built the collection after watching AI assistants produce dozens of low-quality diagrams during a documentation rewrite. The fixed template library removed the back-and-forth editing cycle that typically followed each AI-generated attempt. Developers using Claude Code or Codex report that referencing the template file cut diagram revision time by an estimated 60 percent in internal benchmarks.

The approach matters because diagram quality directly affects how well technical documentation communicates with both engineers and stakeholders. A clean sequence diagram clarifies integration points; a cluttered one obscures them. This repo makes the clean version the default for AI-assisted documentation workflows.

Public APIs remains the definitive free API catalog

public-apis/public-apis is a community-maintained directory of free APIs for use in software projects, currently sitting at 474,008 stars. First published in 2016 by Brian Phung, the repository has grown into the most-starred open-source API catalog on GitHub, exceeding the star counts of most framework and language projects combined.

The directory organizes APIs by category: Animation, Authentication, Crypto, Developer Tools, Finance, Health, Music, Search, Weather, and dozens more. Each entry includes the API name, a one-line description, a link to the documentation, and a tag indicating whether the service is free, freemium, or requires an account. As of August 2026, the list contained over 2,800 individual API entries.

The project uses a simple Markdown file as its source of truth. Contributors submit pull requests with new APIs, and maintainers review them for accuracy before merging. This lightweight governance model has kept the catalog accurate through three major updates to how API monetization works, including the shift toward freemium tiers that many earlier directories mishandled.

According to the GitHub repository, the project receives roughly 200 new contribution pull requests per month. The most active contributors have maintained entries for over five years, which explains why the catalog remains reliable enough for enterprise procurement teams to reference when evaluating third-party integrations.

Modular's Mojo language targets Python developers who need raw performance

modular/modular is the official repository for the Modular Platform, which includes the Mojo programming language designed to run Python code at near-C speed. The project accumulated 29,421 stars by offering a drop-in Python compatibility layer that unlocks GPU acceleration, SIMD parallelism, and zero-cost abstractions without requiring developers to abandon existing Python codebases.

Mojo compiles to LLVM bytecode, the same infrastructure used by Clang and Swift. This means a Python function annotated with Mojo's type system can execute at machine-code speed while retaining Python's syntax and package ecosystem. The Modular team calls this "the best of both worlds," though the practical reality is more nuanced: well-annotated code runs fast, unannotated code falls back to the Python interpreter.

According to Modular's own benchmarks published on their developer documentation site, a matrix multiplication kernel written in Mojo ran 300 times faster than the equivalent pure-Python implementation and within 5 percent of a hand-optimized C version. Those numbers matter for AI inference workloads, where Python dominates development but speed constraints push teams toward CUDA or Triton.

The repository also ships the Modular SDK, CLI tooling, and a growing set of example projects. Development activity is high: the main branch sees multiple commits per day, and the issue tracker reflects an active user base reporting bugs and requesting features for the upcoming Mojo 2.0 release scheduled for late 2026.

Holehe performs email reconnaissance across social platforms

megadose/holehe is a Python tool that checks whether an email address is registered on platforms like Twitter, Instagram, LinkedIn, and Doxed, returning connection status for each service. The project holds 14,568 stars and has become a standard utility in the OSINT (Open-Source Intelligence) toolkit used by security researchers and penetration testers.

The tool works by sending registration-check requests to each supported platform and parsing the response to determine whether the email is associated with an existing account. It does not crack passwords or extract personal data. It answers a single question: is this email address in use on service X?

According to the GitHub repository, holehe supports over 120 platforms out of the box. The project uses asynchronous HTTP requests, which means a full scan across all supported services completes in under 30 seconds on a standard broadband connection. Rate limiting is respected through configurable delays between requests.

Security researchers at Trend Micro have cited holehe in threat intelligence reports as a common initial-reconnaissance tool observed in targeted phishing campaigns. The tool's open-source nature means anyone can inspect the code for privacy concerns, which distinguishes it from commercial alternatives that operate as black boxes.

Semantica builds graph-native infrastructure for accountable AI systems

semantica-agi/semantica is a graph-native infrastructure layer that stores context, enforces accountability, and provides audit trails for AI decision-making pipelines. The project reached 11,570 stars by positioning itself as the middleware between LLM applications and the compliance requirements that enterprise deployments demand.

The core insight behind Semantica is that traditional vector databases store embeddings in isolation, making it difficult to trace why an AI system produced a specific output. Semantica instead represents data as a property graph where nodes are entities, edges are relationships, and each node carries a lineage record showing its source, timestamp, and confidence score.

According to the project documentation, the system integrates with LangChain, LlamaIndex, and direct OpenAI/Anthropic API calls. It stores every prompt-response pair alongside the retrieved context that informed the model's answer, creating a complete decision trace that auditors can review without reconstructing the retrieval pipeline from scratch.

The project is written in Python and requires a PostgreSQL database with the pg_graph extension. Setup takes approximately 20 minutes according to the installation guide. The team behind Semantica comes from regulated-industry backgrounds in fintech and healthcare, which explains the emphasis on compliance features like GDPR right-to-erasure support and automated data-retention policies.

Needle runs a 14MB foundation model on phones, wearables, and robots

cactus-compute/needle is a 14-megabyte foundation model designed to run locally on resource-constrained devices including smartphones, smartwatches, and embedded robots. At 9,937 stars, the project demonstrates how far efficient model compression has advanced: a model small enough to fit on an old USB drive can perform reasoning tasks that previously required cloud API calls.

The Needle model is built on a distilled transformer architecture with quantized weights in the INT4 format. It supports text completion, basic code generation, and instruction following. The developers achieved the small footprint by combining knowledge distillation from a larger teacher model with aggressive pruning of attention heads that contributed least to downstream task performance.

According to the GitHub repository, Needle runs on ARM Cortex-M microcontrollers, Raspberry Pi 4, and Android devices with as little as 64MB of RAM. Latency measurements show responses in 800 milliseconds to 2 seconds depending on hardware, compared to the 2-5 second round-trip typical of cloud-based LLM APIs for equivalent prompt lengths.

The project matters for edge-AI applications where latency, privacy, or connectivity constraints make cloud inference impractical. Medical devices in operating rooms, industrial robots on factory floors, and offline field tools for disaster response all benefit from having reasoning capability that does not depend on cellular or Wi-Fi availability. Running inference locally also avoids the background telemetry and cloud dependencies that performance optimization guides identify as major system overhead sources.

Switchyard routes LLM traffic across models to optimize cost and quality

NVIDIA-NeMo/Switchyard is a routing layer that directs LLM application requests across multiple models and providers while preserving native OpenAI and Anthropic API compatibility. The project, with 2,636 stars, solves the growing problem of multi-provider AI cost management: as pricing shifts between GPT-4o, Claude 3.5 Sonnet, and open-weight alternatives, applications need dynamic routing that responds to both price changes and quality requirements in real time.

Switchyard implements a decision engine that evaluates each incoming request against configurable rules. Simple queries route to cheaper models; complex reasoning tasks route to premium models; rate-limit breaches trigger automatic failover to backup providers. The routing layer sits between the application and the LLM API, intercepting requests before they reach any provider.

According to the project README, Switchyard maintains a transparent proxy that logs every routing decision with the model used, tokens consumed, latency observed, and cost incurred. This audit log enables finance teams to reconcile AI spending against actual usage patterns rather than relying on provider invoices alone.

NVIDIA's NeMo team built Switchyard as part of a broader effort to make enterprise AI deployments more resilient and cost-efficient. The project ships Docker containers, Kubernetes manifests, and a REST management API. Teams running multiple LLM contracts through Switchyard report average cost reductions of 35 to 50 percent without measurable quality degradation on standard benchmark suites.

Shockwave is a local-first notes app with a built-in coding agent

stephengpope/shockwave is a file-based note-taking application where all content lives as plain Markdown files in a folder the user owns, shipped with a coding agent integrated directly into the editor. At 188 stars, the project is early-stage but represents a trend toward local-first tools that refuse to lock users into proprietary ecosystems.

The app stores notes as individual .md files in a specified directory. There is no database, no cloud sync requirement, and no proprietary file format. Opening the folder in any text editor produces the same content that Shockwave displays. The built-in coding agent can read, modify, and generate Markdown files using local LLM inference through OpenRouter or local Ollama instances.

According to the repository description, the agent integration distinguishes Shockwave from competitors like Obsidian and Logseq, which require third-party plugins for AI assistance. Here the agent is the first-class citizen: users can highlight a passage, ask the agent to summarize it, rewrite it in a different tone, or expand it into a full section, and the changes write directly back to the source Markdown file.

The project uses Electron for the desktop shell and is available on Windows, macOS, and Linux. Development is active but slow, with updates released monthly rather than weekly. The author has signaled that the next major release will add bi-directional sync with Git repositories, enabling version history and collaborative editing without leaving the local-first paradigm.

Lumina packages a full agentic desktop harness for local inference

Bino5150/lumina is a desktop application that wraps local AI inference into a full agentic workflow environment, designed to run efficiently on consumer-grade hardware. With 157 stars, it is the youngest project on this list, but its design philosophy aligns with a broader movement toward self-hosted AI assistants that do not leak conversation data to third-party clouds.

Lumina provides a graphical interface for managing multiple AI agents, each configured with different system prompts, model endpoints, and context windows. The default configuration uses Ollama for local model serving and supports running Llama 3.2, Phi-4, and Mistral variants on a single GPU with 8GB of VRAM. The application includes a built-in terminal, file browser, and web request tool so agents can perform multi-step tasks without leaving the desktop.

According to the GitHub repository, Lumina distinguishes itself from competing local agent projects through its resource management layer. The application monitors GPU memory, CPU temperature, and system RAM in real time, throttling agent activity when hardware utilization exceeds safe thresholds. This prevents the kind of thermal throttling that makes local inference frustrating on laptop hardware.

The project is written in Python and distributed as a PyInstaller bundle for easy installation. The author, Bino5150, has been shipping updates since June 2026 and maintains an active Discord channel where users share configuration profiles and report hardware compatibility issues across different GPU generations.

Imagine CLI turns the terminal into an AI image generator

AhmedAburady/imagine-cli is a command-line tool that generates AI images through a single YAML configuration file, eliminating environment variable setup and boilerplate code. The project sits at 50 stars and targets developers who want image generation capabilities inside terminal workflows, CI/CD pipelines, and automation scripts without writing Python glue code.

The tool abstracts away the complexity of API authentication, request formatting, and response parsing. Users create a single YAML file specifying the model, dimensions, prompt, and output path, then run one command to generate the image. Supported providers include Stable Diffusion XL, DALL-E 3, and Midjourney via unofficial endpoints. The project handles pagination, retry logic, and rate-limit backoff automatically.

According to the repository, the YAML-first design choice was intentional. Most existing CLI image generators require command-line flags for every parameter, which becomes unwieldy for batch generation tasks where the same settings repeat across hundreds of images. A config file can be version-controlled, shared across team members, and modified without memorizing flag syntax.

The project is in early development with an active contributor base. AhmedAburady has published integration guides for popular terminal emulators and shell aliases that let users generate images with a two-key shortcut from any working directory.

Why these repos matter right now

The repositories highlighted here reflect three concurrent shifts in how developers build with AI. The first is infrastructure maturation: projects like OpenViking and Semantica are moving agent development from experimental notebooks to production systems with memory, auditability, and compliance features. The second is efficiency pressure: Needle and Switchyard show that cost optimization is no longer optional. Cloud API bills are scaling faster than most engineering budgets anticipate, and the tools that help teams do more with fewer tokens are becoming table stakes. The third is local-first sovereignty: Shockwave, Lumina, and Imagine CLI all share a design principle that data should remain under user control, whether that means storing notes as plain files or running inference on hardware the user owns.

Together these trends suggest that the next phase of AI tooling will be defined less by model capability and more by how efficiently that capability can be embedded into existing workflows without creating new vendor dependencies. If you want to revisit how this space looked two weeks earlier, our coverage of the previous GitHub repo roundup covers overlapping themes around Claude plugins, local AI, and token cost.

Conclusion

Eleven of the thirteen repos featured in The Next New Thing's August 21 episode are new additions to TekMag's coverage, spanning agent memory systems, editorial diagram templates, local-first infrastructure, and cost-aware routing. The common thread is practical utility over speculative capability.

Frequently Asked Questions

{"@context": "https://schema.org", "@type": "FAQPage", "mainEntity": [{"@type": "Question", "name": "What is the most-starred repo on this list?", "acceptedAnswer": {"@type": "Answer", "text": "public-apis/public-apis leads with 474,008 stars, making it the most-starred API catalog on GitHub and one of the oldest continuously maintained open-source directories."}}, {"@type": "Question", "name": "Can OpenViking work with existing RAG pipelines?", "acceptedAnswer": {"@type": "Answer", "text": "Yes. OpenViking exposes a REST API compatible with standard LangChain and LlamaIndex integrations, allowing teams to migrate existing RAG setups without rewriting their retrieval logic."}}, {"@type": "Question", "name": "Does Switchyard actually reduce LLM costs?", "acceptedAnswer": {"@type": "Answer", "text": "According to the NVIDIA NeMo team, teams routing traffic through Switchyard report average cost reductions of 35 to 50 percent while maintaining quality on standard evaluation benchmarks."}}, {"@type": "Question", "name": "Is Holehe legal to use for security research?", "acceptedAnswer": {"@type": "Answer", "text": "Holehe only checks whether an email is associated with a public account on supported platforms. It does not access private data. Security researchers and penetration testers use it within the scope of authorized assessments."}}, {"@type": "Question", "name": "What hardware do I need to run Needle locally?", "acceptedAnswer": {"@type": "Answer", "text": "Needle runs on devices with as little as 64MB of RAM, including ARM Cortex-M microcontrollers, Raspberry Pi 4, and Android smartphones. No dedicated GPU is required."}}]}

What is the most-starred repo on this list?

public-apis/public-apis leads with 474,008 stars, making it the most-starred API catalog on GitHub and one of the oldest continuously maintained open-source directories.

Can OpenViking work with existing RAG pipelines?

Yes. OpenViking exposes a REST API compatible with standard LangChain and LlamaIndex integrations, allowing teams to migrate existing RAG setups without rewriting their retrieval logic.

Does Switchyard actually reduce LLM costs?

According to the NVIDIA NeMo team, teams routing traffic through Switchyard report average cost reductions of 35 to 50 percent while maintaining quality on standard evaluation benchmarks.

Is Holehe legal to use for security research?

Holehe only checks whether an email is associated with a public account on supported platforms. It does not access private data. Security researchers and penetration testers use it within the scope of authorized assessments.

What hardware do I need to run Needle locally?

Needle runs on devices with as little as 64MB of RAM, including ARM Cortex-M microcontrollers, Raspberry Pi 4, and Android smartphones. No dedicated GPU is required.

References

1. The Next New Thing, "Top 10 GitHub: AI videos, gorgeous diagrams, token savings and more," YouTube, published August 21, 2026. Watch the episode.
2. volcengine, "OpenViking: Self-evolving Context Database for AI Agents." GitHub repository, 34,849 stars. View repository.
3. cathrynlavery, "Diagram Design: 38 Editorial Diagram Types for AI Coding Assistants." GitHub repository, 29,190 stars. View repository.
4. public-apis, "A collective list of free APIs." GitHub repository, 474,008 stars. View repository.
5. Modular, "Mojo Programming Language." Official documentation. modular.com/mojo.
6. megadose, "Holehe: Email osint tool." GitHub repository, 14,568 stars. View repository.
7. semantica-agi, "Semantica: Graph-Native Infrastructure for Context and Accountable AI Systems." GitHub repository, 11,570 stars. View repository.
8. cactus-compute, "Needle: 14MB Foundation Model for Tiny Devices." GitHub repository, 9,937 stars. View repository.
9. NVIDIA-NeMo, "Switchyard: LLM Traffic Routing Across Models and Providers." GitHub repository, 2,636 stars. View repository.
10. stephengpope, "Shockwave: Local-first notes with a built-in coding agent." GitHub repository, 188 stars. View repository.
11. Bino5150, "Lumina: Agentic Desktop Harness for Local Inference." GitHub repository, 157 stars. View repository.
12. AhmedAburady, "Imagine CLI: Terminal AI Image Generation." GitHub repository, 50 stars. View repository.
13. TekMag, "Top 10 Repos: 2,282 Claude plugins, free codex, & 8 more," published August 31, 2026. Read the previous article.