RepoRadar helps AI developers, technical founders, and engineering leads answer three questions: What changed? What should I try? Which option fits this workflow?
Saved tools, searches, risks, movers, and newly reviewed candidates appear here first.
Check the newest retained change window when you are ready. Nothing is loaded until you ask.
Open the full change history →The latest retained evidence-backed picks, selected from editorial or hands-on review. Each card shows its actual verification stage and opens the RepoRadar record first.
pydantic/pydantic-ai is an MIT-licensed Python agent framework from the Pydantic team that gives developers a type-safe, model-agnostic, FastAPI-style way to build production LLM…
Apache-2.0 licensed.
OpenAI Agents SDK for JavaScript/TypeScript is an MIT-licensed, provider-agnostic framework for building multi-agent and voice-agent workflows with tools, MCP support, handoffs…
Fission-AI/OpenSpec is an MIT spec-driven development framework and CLI for AI coding assistants that turns vague requests into reviewable proposals, plans, task artifacts, and…
microsoft/presidio is an MIT-licensed, context-aware PII de-identification SDK from Microsoft that detects and anonymizes sensitive entities in text and images and ships pluggable…
MIT Microsoft cross-platform ML inference and training accelerator with ONNX standard support + hardware accelerators across CPU / GPU / NPU; 21,042* at verify time, last commit…
Repeat-visitor view: new signals, score movement, risk-label changes, and fresh try-now picks from the 24 hours to 2026-09-13.
Tools, repos, agents, apps, and frameworks added in the 24 hours to 2026-09-13.
New or newly tracked model-related signals in the 24 hours to 2026-09-13.
No change detected in the 24 hours to 2026-09-13.
No change detected in the 24 hours to 2026-09-13.
No change detected in the 24 hours to 2026-09-13.
Newly added or newly promoted try-now verdicts in the 24 hours to 2026-09-13.
OpenAI's companion repository for its announced finite-time-blowup results for the Navier-Stokes and Euler equations. It publishes the Lean 4 formalizations of both papers together
An inference runtime that streams mixture-of-experts weights from NVMe SSD so large sparse models can run on consumer GPUs with limited VRAM. Ships configs, quantization and servin
Saivineeth147/lora-speedrun is an MIT-licensed fine-tuning speedrun harness (148 stars, 10 forks, 4 contributors returned by the GitHub contributors API, created 2026-07-18, pushed
Commerce Agent Bench is the renamed Accio-org/RealReplicaBench — the Accio team's (Alibaba International) benchmark for long-horizon agents in reproducible replicas of commerce sof
A ready-to-install orchestration layer for the OpenAI Codex CLI: an 'Astra' coordinating profile plus 'Luna' subagent profiles, with install and uninstall scripts, config profiles,
Reviewed decision records from this run. Each passed the publication gate and links to its evidence and verification basis. Rotates weekly so returning visitors see fresh picks.
Why it's featured: it cleared our evidence and quality gates with the strongest overall signal in this run — the pick we’d point a newcomer to first.
Open decision record →Tested in a bounded workflowGitHub source ↗ for langchain-ai/langchainMomentum is a research lead, not a recommendation or editorial verification. Trending reflects engagement velocity and recency; each card shows its actual verification stage when one exists.
deepseek-ai/deepseek-harness is a 91.6k-star / 8.3k-fork / 12,293-commit MIT open-source agent harness from DeepSeek built on the Cordis framework…
Research signal · Discovered, not yet verified GitHub source ↗ for deepseek-ai/deepseek-harnessJustVugg/colibri is an Apache-2.0 inference engine written in C with no engine dependencies (26,472 stars, 2,892 forks, 100+ contributors, created…
Research signal · Discovered, not yet verified GitHub source ↗ for colibrì — pure-C inference engine that runs 744B to 2.8T MoE models by treating disk, RAM and VRAM as one memory hierarchycrawl4ai unclecode/crawl4ai is an Apache-2.0 open-source Python library and CLI for LLM-friendly web crawling and scraping that produces clean…
Research signal · Discovered, not yet verified GitHub source ↗ for unclecode/crawl4aiopen-code-review alibaba/open-code-review is an Apache-2.0 open-source AI-powered code-review CLI that originated as Alibaba Group's internal…
Research signal · Discovered, not yet verified GitHub source ↗ for alibaba/open-code-reviewCaveman JuliusBrussee/caveman is an MIT-licensed open-source skill/plugin for Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and 30+…
Research signal · Discovered, not yet verified GitHub source ↗ for JuliusBrussee/cavemanearendil-works/pi is an MIT agent toolkit that bundles a unified multi-provider LLM API, agent runtime, interactive coding-agent CLI, terminal UI…
Tested in a bounded workflow GitHub source ↗ for earendil-works/piSource-backed candidates chosen for further research. This is an editorial watchlist, not a measured growth ranking; each card shows its current catalog score and retained verification stage.
LangChain is a Python/JavaScript agent-engineering framework for composing model calls, tools, retrieval, structured outputs, multi-agent flows, and provider integrations.
An AI-powered job search system layered on Anthropic's Claude Code CLI.
N8n is a Fair-code workflow automation platform with native AI capabilities.
LocalAI is a self-hosted OpenAI-compatible inference server for running LLMs, embeddings, image/audio generation, speech, rerankers, and vision-style workloads locally.
Dify is a low-code platform for building LLM apps: chatbots, agent workflows, RAG knowledge bases, prompt orchestration, deployment, and operational monitoring.
LobeHub is a web-based agent operations hub for creating specialized AI agents, scheduling them as always-on jobs, attaching knowledge bases, and reviewing their run reports.
The latest tools entering the research catalog. Filtered by recency and catalog score for triage; inclusion is not editorial verification.
OpenAI's companion repository for its announced finite-time-blowup results for the Navier-Stokes and Euler equations. It publishes the Lean 4 formalizations of both papers...
An inference runtime that streams mixture-of-experts weights from NVMe SSD so large sparse models can run on consumer GPUs with limited VRAM. Ships configs, quantization and...
A ready-to-install orchestration layer for the OpenAI Codex CLI: an 'Astra' coordinating profile plus 'Luna' subagent profiles, with install and uninstall scripts, config...
A fast, GPU-rendered, keyboard-driven IDE built in Rust, featuring code editing, terminals, debugging, language intelligence, and the Rune Agent, an AI coding agent shipped as...
A full-stack open-source build of a palm-sized RL-trained biped robot: mechanical design, electronics, firmware, training pipeline, and a step-by-step tutorial that takes you...
GPU kernels and sampler implementations for the DeepSeek Sparse Attention TopK selection, released by the DeepSeek-AI organization with reference implementations so other...
The homepage separates reviewed decision surfaces from clearly labeled research signals. Search, filters, source links, risk labels, and all 3,393 ranked cards are available in the Research Database, with 307 MCP records catalogued separately.
3,700 open-source: 3,393 Research Database · 307 MCP; plus 7 commercial.
Fresh and recently refreshed items in the public dataset.
Items with a practical verdict worth evaluating now.
Top evidence-linked useful signals.
RepoRadar is independently run. Tips help cover hosting, sources, and the time it takes to keep this useful. Never ranking. Never coverage. Never results.
V1 scoring prioritizes practical usefulness: usefulness 35%, novelty 18%, momentum 14%, maturity 10%, open-source/build quality 7%, evidence 6%, commercial/workflow potential 6%, setup ease 4%.
Gold is score ≥ 7.5 or the top 10% of eligible items with score ≥ 7.0. Silver starts at 6.5, Bronze at 5.25, and Low Signal catches low-value or low-signal items.
Gold requires Medium/High confidence, usable/official/cross-verified evidence or better, clear public descriptions, no excluded category, no weak or unclear AI relevance, and no fallback, triage-only, weak-evidence, low-signal, or low-transparency flags.
Risk remains a separate label and should not heavily reduce score unless severe. Transparency issues apply small penalties; hard exclusions still block publication.