New apps and open-source projects for AI builders.
Vercel UI SDK provides composable interaction behaviors for building interfaces across JavaScript frameworks.
- Interaction primitives. The shared core handles focus, keyboard navigation, selection, dismissal, drag and drop, and associated ARIA attributes. Source ↗
- Framework support. The beta ships a core package and separate adapters for React, Vue, Svelte, Solid and Web Components. Source ↗
Scheduled · awaiting first completed check
LiteLLM provides a Python SDK and self-hosted gateway for routing requests to model providers through a common API.
- Routing. The router supports retries, fallbacks and strategies such as least-busy or weighted routing. Source ↗
- Access controls. The proxy can issue virtual keys tied to allowed models, budgets and rate limits. The documented key setup uses PostgreSQL. Source ↗
Scheduled · awaiting first completed check
AnythingLLM packages document retrieval, chat and model connections into self-hosted workspaces.
- Desktop and server. Desktop runs as a single-user app. The Docker deployment supports multiple users and roles. Source ↗
- Document isolation. A workspace controls which embedded documents its chats can draw from. Source ↗
Scheduled · awaiting first completed check
Unsloth provides tools for running, fine-tuning and exporting open-weight models on local hardware.
- Fine-tuning. The toolkit supports LoRA, 4-bit QLoRA and full fine-tuning, which needs more compute and memory. Source ↗
- Export. Studio exports merged weights, adapter-only weights that require the original model, or GGUF files for local runtimes. Source ↗
Scheduled · awaiting first completed check
Open WebUI is a self-hosted interface for chatting with local models or OpenAI-compatible endpoints.
- Model backends. The application connects to Ollama or OpenAI-compatible servers, so inference runs on the configured backend. Source ↗
- Document retrieval. Its built-in retrieval layer can attach document collections to a chat and return source citations. Source ↗
Scheduled · awaiting first completed check
AIHOT is a self-hosted news engine that turns configured sources and selection criteria into a topical site and digest.
- Selection. An initial filter and two independent model scores decide which items clear a source-tier threshold. Prompts and cutoffs are configurable. Source ↗
- Event grouping. The engine groups coverage of the same event and weights recency and distinct sources in the trending list. Source ↗
Scheduled · awaiting first completed check
E2B provides hosted, isolated Linux sandboxes for running application and agent workloads.
- Sandbox controls. Python and JavaScript SDKs expose command execution and file operations in a sandbox. Source ↗
- Templates. Custom templates define the packages and files a sandbox starts with, alongside build steps and a startup command. Source ↗
Scheduled · awaiting first completed check
llama.cpp is a C and C++ runtime for running language and multimodal models on local hardware.
- Weights. The quantization tool converts higher-precision GGUF weights to lower-precision GGUF files, reducing size with a possible accuracy cost. Source ↗
- Hardware. The runtime supports CPU execution and accelerator backends including Metal, CUDA and Vulkan. Source ↗
Scheduled · awaiting first completed check
Gradio wraps Python model functions and workflows in interactive web demos.
- UI composition. Interface wraps a function with typed inputs and outputs, while Blocks exposes layout, event wiring and multi-step flows. Source ↗
- Queueing. Event handlers can share a queue by concurrency ID and enforce a concurrency limit for scarce resources such as a GPU. Source ↗
Scheduled · awaiting first completed check
Ollama runs and serves local models and connects to separately hosted cloud models.
- Authentication. The local API does not require authentication by default. The hosted endpoint uses an API key, and cloud models accessed through the local server require sign-in. Source ↗
- Model imports. A Modelfile can import GGUF or supported Safetensors models. Ollama does not quantize GGUF files during import. Source ↗
Scheduled · awaiting first completed check
Repomix packages a repository into a single file for code review and model context.
- Packaging. The CLI can pack local or remote repositories and produce XML, Markdown, JSON or plain text with a file tree and file contents. Source ↗
- Context reduction. Optional Tree-sitter compression retains structural declarations while dropping implementation detail in supported languages. Source ↗
Scheduled · awaiting first completed check
Streamlit turns Python scripts into interactive web apps for data and model workflows.
- Execution model. Widget interaction ordinarily reruns the script from top to bottom. Per-session state persists across reruns. Source ↗
- Caching. Cache data stores returned values and gives callers copies. Cache resource can share live objects such as connections or models. Source ↗
Scheduled · awaiting first completed check
New Term Radar collects early signals about emerging AI names and produces a daily term report.
- Source signals. It reads Hugging Face trends, fast-growing GitHub repositories and sitemap changes, with optional feeds such as X and Search Console. Source ↗
- Access and costs. The Python package is MIT-licensed. Public feeds can run without a paid key, while X and keyword research use separate service credentials. Source ↗
Scheduled · awaiting first completed check
Replicate Cog packages Python models and their dependencies into containers with a prediction API.
- Build configuration. The cog.yaml file declares the Python environment, dependencies and optional GPU support for the image. Source ↗
- API contract. Typed Python runner methods define prediction inputs and outputs, which Cog exposes through a generated OpenAPI schema. Source ↗
Scheduled · awaiting first completed check
PRINTFILM is a self-hosted production workspace for AI short videos and episodic illustrated dramas.
- Production workflow. Short videos move through editable script, storyboard, reference-image generation, clip generation and assembly, with a review stop between stages. Source ↗
- Episodic assets. Drama projects keep reusable characters, scenes and props in an asset library shared across episodes. Source ↗
Scheduled · awaiting first completed check
x402 is an HTTP payment protocol for charging for a request using machine-readable terms and signed payment authorization.
- HTTP exchange. A server sends payment requirements with HTTP 402. The client retries with a signed payment payload and can receive settlement details in a response header. Source ↗
- Payment schemes. The exact scheme charges a fixed amount, while upto authorizes a maximum for metered usage. Support depends on the scheme and network pair. Source ↗
Scheduled · awaiting first completed check
Open Intelligent UI is a reference chat app that streams model output as interactive components using the OpenUI language.
- Rendering. OpenUI Lang 1.0 streams into a component registry that includes forms, charts and maps, with variables that update as users interact. Source ↗
- Model access. The Node 24 app supports a server-side OpenUI Cloud key, a separately configured OpenAI-compatible API, or a local Ollama endpoint. Source ↗
Scheduled · awaiting first completed check
Solid Glass is a CSS theme and shadcn registry for building AI application interfaces.
- Integration. The stylesheet targets existing shadcn/ui slots and classes. Its documented setup uses React 19, Tailwind CSS 4 and Base UI. Source ↗
- Components. The registry includes a composer, tool-step view, diff view and context indicator alongside the theme. Source ↗
Scheduled · awaiting first completed check