multimodal

Mintplex-Labs/anything-llm

A local-first LLM workspace that turns documents into working agents — AnythingLLM bundles RAG chat, multi-user support, and configurable AI agents into one self-hosted app, aimed at people who want to run document-grounded chat and agentic workflows locally instead of relying on hosted AI services. It targets self-hosters and teams needing private, multimodal, agent-capable LLM tooling without complex setup.

bojieli/ai-agent-book

An open-source book on AI Agent design with 109 runnable labs — A free, open 10-chapter book on AI Agent engineering, framed as LLM + context + tools, by Li Bojie. It ships 109 hands-on experiments, PDF/EPUB downloads, an online reader, and translations into 15 languages, aimed at engineers building real agent systems.

bytedance/UI-TARS-desktop

The Open-Source Multimodal AI Agent Stack: Connecting Cutting-Edge AI Models and Agent Infra

screenpipe/screenpipe

Screenpipe records your work screen to surface automation targets — Screenpipe continuously captures screen and audio on company computers, then maps workflows to find tasks worth automating and feeds context to AI agents. Backed by YC S26, it's built local-first and privacy-conscious, aimed at teams wanting to understand and automate real work patterns rather than guess at them.

rerun-io/rerun

A columnar data layer built for multimodal robotics logging — Rerun ingests multi-rate robotics data—images, point clouds, transforms, time series, video—from logs, sim, or MCAP/LeRobot formats, then lets you visualize it synchronously in a built-in viewer, query it as dataframes or SQL, and stream it into training. Built in Rust with Python, Rust, and C++ SDKs, it targets robotics and physical AI teams handling heterogeneous sensor data.

apache/seatunnel

SeaTunnel extends data integration to multimodal and LLM workloads — Apache SeaTunnel syncs huge data volumes across 160+ connectors, now covering video, images, and unstructured text alongside structured data for embeddings and LLM pipelines. It runs on Zeta, Flink, or Spark engines, supporting batch, streaming, and CDC with a distributed snapshot algorithm for consistency. Useful for teams needing flexible, multi-engine data pipelines.

ARahim3/DigUp

Search your computer by what's in your files: photos, PDFs, documents, recordings, videos and code. On-device on Mac, Windows and Linux, built on Google DeepMind's EmbeddingGemma 2.

sgl-project/sglang-omni

A SGLang spin-off dedicated to serving speech and omni models — Built on the SGLang project, SGLang-Omni focuses on serving TTS, ASR and unified multimodal models with OpenAI-compatible APIs, streaming and distributed inference on CUDA. Useful for teams who need production-grade speech and multimodal serving without building custom inference stacks.

uclanlp/OpenVLHarness

Official Code for OpenVLHarness: Real-World Visual Agentic Tool‑Use with Multimodal Memory

OmniJev/Flide

Exact state forking speeds decision model inference 9.5x over vLLM — Flide is inference infrastructure built for Jev-style decision models. Its exact state forking lets it serve the Jev /v1/systemone API up to 9.5 times faster than vLLM and SGLang, while guaranteeing each answer matches what the model would give for that question alone. Useful for teams running LLM-based decision or judging pipelines who need both speed and deterministic output.