RepoRadar - Useful AI tools, repos, models, and research

OWASP/www-project-agent-memory-guard

Wed, 29 Jul 2026 21:27:23 +0000

OWASP Agent Memory Guard is an Apache-2.0 Python package for defending AI-agent memory against poisoning, tool abuse, privilege escalation, and excessive autonomy. Current primary checks show an active OWASP repository, raw Apache-2.0 LICENSE.md, pyproject metadata for package `agent-memory-guard` v0.3.0, and matching PyPI repository metadata. RepoRadar tier: Gold.

FluxVLA/FluxVLA

Wed, 29 Jul 2026 21:27:23 +0000

FluxVLA is an Apache-2.0 engineering platform for vision-language-action work, covering data, model serving, and real-robot deployment workflows. Current GitHub metadata shows active maintenance, 577 stars, 59 forks, a 2026-07-29 push, and raw Apache-2.0 LICENSE verification; the README points to Hugging Face assets and documentation for the VLA engine. RepoRadar tier: Gold.

Bestpart-Irene/secondsign-core

Wed, 29 Jul 2026 21:27:23 +0000

SecondSign Core is an Apache-2.0 Python project for runtime authorization around financial AI agents. The README describes structured intent, deterministic policy, human approval, execute-once gateways, and tamper-evident receipts; current checks verified an active GitHub repo, raw Apache-2.0 LICENSE, pyproject license metadata, and a matching PyPI package repository. RepoRadar tier: Silver.

loveramarois-byte/council-lab

Wed, 29 Jul 2026 19:26:54 +0000

Council Lab is an Apache-2.0 local-first workspace for running a structured five-seat AI deliberation: four sequential agent viewpoints plus a final summarizer after a user confirmation point. The current README documents visible debate turns, persisted checkpoints, resumable runs, provider/model snapshots, SQLite migrations, local credential handling, and exportable reports. RepoRadar tier: Gold.

microsoft/Fara1.5-27B

Wed, 29 Jul 2026 19:26:54 +0000

Fara1.5-27B is Microsoft's MIT-licensed 27B multimodal computer-use model on Hugging Face. The model card says it observes browser screenshots and emits structured tool calls such as click, type, scroll, URL visit, and web search; Hugging Face metadata shows a current 2026-07-27 model update, 1,543 downloads, 195 likes, safetensors, and transformers support. RepoRadar tier: Gold.

Scaling GUI Agents with Visual State Transitions

Wed, 29 Jul 2026 19:26:54 +0000

This arXiv paper introduces State Transition Pretraining for GUI agents: a multimodal model learns from visual before/after interface states by jointly predicting actions and next states, then is fine-tuned on task trajectories. The arXiv API lists the paper as version 1 submitted 2026-07-27 and the abstract reports improvements on desktop and mobile GUI-agent evals. RepoRadar tier: Gold.

ggml-org/llama.cpp

Wed, 29 Jul 2026 18:27:47 +0000

llama.cpp is an MIT-licensed C/C++ runtime for local LLM inference with GGUF models, server builds, Docker packaging, and a very large maintainer/user footprint. Current GitHub metadata shows the canonical repo under ggml-org, active maintenance, and a raw MIT license. RepoRadar tier: Gold.

axolotl-ai-cloud/axolotl

Wed, 29 Jul 2026 18:27:47 +0000

Axolotl is an Apache-2.0 framework for fine-tuning language models. The current README describes it as a free and open-source LLM fine-tuning framework, while GitHub metadata and the raw LICENSE show active maintenance and Apache-2.0 terms. RepoRadar tier: Gold.

NVIDIA/Model-Optimizer

Wed, 29 Jul 2026 18:27:47 +0000

NVIDIA Model Optimizer is an Apache-2.0 library for quantization, pruning, distillation, neural architecture search, speculative decoding, and sparsity. The current README points to documentation and PyPI packaging, and the raw repository license is Apache-2.0. RepoRadar tier: Gold.

enactic/openarm

Wed, 29 Jul 2026 18:27:47 +0000

OpenArm is an Apache-2.0 humanoid-arm platform for physical-AI research, teleoperation, imitation learning, simulation, and real-world data collection. The README describes a human-scale 7DOF arm and a standardized OpenArm Cell, while GitHub metadata shows active maintenance. RepoRadar tier: Gold.

microsoft/Mage

Wed, 29 Jul 2026 18:27:47 +0000

Mage is a Microsoft multimodal model-family repository with MIT-licensed Mage-Flow code and README badges that identify Mage-VL under Apache-2.0. GitHub metadata shows current maintenance and the raw repository LICENSE is MIT. RepoRadar tier: Gold.

NVIDIA/garak

Wed, 29 Jul 2026 17:26:59 +0000

Garak is NVIDIA's Apache-2.0 LLM red-team scanner. The current README describes probes for hallucination, data leakage, prompt injection, misinformation, toxicity, jailbreaks, and other unwanted model behaviors, while GitHub metadata shows active maintenance and substantial adoption. RepoRadar tier: Gold.

huggingface/trl

Wed, 29 Jul 2026 17:26:59 +0000

TRL is Hugging Face's Apache-2.0 library for training transformer language models with reinforcement learning and related post-training methods. Current GitHub metadata shows active maintenance, a large contributor/user footprint, and clear positioning as part of the Hugging Face training stack. RepoRadar tier: Gold.

huggingface/diffusers

Wed, 29 Jul 2026 17:26:59 +0000

Diffusers is Hugging Face's Apache-2.0 PyTorch library for working with diffusion models across image, video, and audio generation. GitHub metadata shows very broad adoption, active maintenance, and a raw Apache-2.0 license in the repository. RepoRadar tier: Gold.

OpenDCAI/DataFlow

Wed, 29 Jul 2026 17:26:59 +0000

DataFlow is an Apache-2.0 pipeline project for generating, cleaning, and preparing LLM datasets with model-assisted operators. GitHub metadata shows thousands of stars, recent maintenance, a raw Apache-2.0 license, and a README focused on end-to-end data preparation rather than a thin prompt demo. RepoRadar tier: Gold.

iwe-org/iwe

Wed, 29 Jul 2026 14:27:13 +0000

IWE is an Apache-2.0 Rust workspace that turns local markdown notes into a knowledge graph with editor LSP features, CLI retrieval, and an MCP server so agents can navigate the same note structure without a database or cloud service. RepoRadar tier: Gold.

swellweb/reame

Wed, 29 Jul 2026 14:27:13 +0000

Reame is an MIT-licensed inference server on top of llama.cpp that targets low-cost CPU hardware first, reuses repeated context across similar requests, and documents measured speed gains on free-tier ARM boxes instead of assuming every useful local deployment starts with a GPU. RepoRadar tier: Gold.

lakeraai/pint-benchmark

Wed, 29 Jul 2026 14:27:13 +0000

Lakera PINT is a MIT-licensed prompt-injection eval suite that publishes a 4,314-sample test set shape, comparison scores for multiple detection systems, and the surrounding notebooks so teams can pressure-test guardrail models against a broader mix of malicious and lookalike inputs. RepoRadar tier: Silver.

Genesis-Embodied-AI/genesis-world

Wed, 29 Jul 2026 13:28:55 +0000

Genesis World is an Apache-2.0 simulation platform for robotics and embodied AI that combines a Python simulation interface, multi-physics engine, robotics-focused rendering, and a cross-platform compiler targeting CUDA, ROCm, Metal, Vulkan, x86, and ARM64. RepoRadar tier: Gold.

pgvector/pgvector

Wed, 29 Jul 2026 13:28:55 +0000

pgvector is an open PostgreSQL extension for vector similarity search that keeps embeddings and application data inside Postgres while supporting exact and approximate nearest-neighbor search, multiple distance functions, HNSW, and IVFFlat indexing. RepoRadar tier: Gold.

espnet/espnet

Wed, 29 Jul 2026 13:28:55 +0000

ESPnet is an Apache-2.0 end-to-end speech-processing toolkit that covers speech recognition, text-to-speech, speech translation, enhancement, diarization, voice conversion, and streaming or offline workflows in one actively maintained research and engineering stack. RepoRadar tier: Gold.

risa-labs-inc/BossConsole

Wed, 29 Jul 2026 13:28:55 +0000

BossConsole is an Apache-2.0 desktop harness for AI agents built with Kotlin Multiplatform, combining an embedded browser, shareable terminal, editor, plugin toolbox, and governed MCP tool layer so operators can run agents against real local tools from one native console. RepoRadar tier: Silver.

volcengine/verl

Wed, 29 Jul 2026 12:32:51 +0000

verl is an Apache-2.0 post-training framework for large language models that packages PPO, GRPO, hybrid-controller dataflows, and integrations with vLLM, FSDP, Megatron-LM, and SGLang into a real training stack teams can build on instead of stitching together one-off RL scripts. RepoRadar tier: Gold.

newton-physics/newton

Wed, 29 Jul 2026 12:32:51 +0000

Newton is an Apache-2.0 GPU-accelerated physics simulation engine for robotics work, built on NVIDIA Warp with MuJoCo Warp integration, OpenUSD support, differentiability, and a real installation path for researchers who need faster simulation loops than legacy CPU-heavy stacks usually provide. RepoRadar tier: Gold.

OpenRLHF/OpenRLHF

Wed, 29 Jul 2026 12:32:51 +0000

OpenRLHF is an Apache-2.0 framework for RLHF and agent-oriented reinforcement learning built around Ray, vLLM, DeepSpeed-style distributed training, and recent backend integrations, giving teams a practical open stack for scalable post-training instead of a collection of disconnected scripts. RepoRadar tier: Gold.

google/langextract

Wed, 29 Jul 2026 11:37:30 +0000

LangExtract is an Apache-2.0 Python library for turning long unstructured text into structured data with exact source grounding, schema control, and an interactive viewer for inspecting what the model pulled from each passage. RepoRadar tier: Gold.

containers/kubernetes-mcp-server

Wed, 29 Jul 2026 11:37:30 +0000

Kubernetes MCP Server is an Apache-2.0 MCP server for Kubernetes and OpenShift that exposes cluster resources, pod operations, and generic CRUD actions through a maintained package for agent-facing platform workflows. RepoRadar tier: Gold.

mark3labs/mcp-go

Wed, 29 Jul 2026 11:37:30 +0000

mcp-go is a MIT Go SDK for building MCP clients and servers with typed support for prompts, resources, tools, transports, and examples, making it a practical foundation for Go-based MCP integrations. RepoRadar tier: Gold.

traceloop/openllmetry

Wed, 29 Jul 2026 11:37:30 +0000

OpenLLMetry is an Apache-2.0 observability layer for LLM applications built on OpenTelemetry, giving teams tracing, evaluation hooks, and provider integrations without inventing a separate telemetry stack from scratch. RepoRadar tier: Gold.

bespokelabsai/curator

Wed, 29 Jul 2026 11:37:30 +0000

Bespoke Curator is an Apache-2.0 toolkit for bulk inference, structured extraction, and synthetic-data curation so teams can build post-training and evaluation datasets without hand-rolling every batch pipeline themselves. RepoRadar tier: Gold.

onecli/onecli

Wed, 29 Jul 2026 10:47:46 +0000

OneCLI is an Apache-2.0 credential gateway with a built-in vault that stores real API secrets once, swaps placeholder credentials at request time, and keeps raw keys out of agent context. RepoRadar tier: Gold.

raketenkater/ggrun

Wed, 29 Jul 2026 10:47:46 +0000

ggrun is a MIT launcher for GGUF models on llama.cpp and ik_llama.cpp that plans tensor placement across mixed GPUs and RAM, exposes an OpenAI-compatible server, and can auto-tune runtime flags for a specific machine. RepoRadar tier: Gold.

Marker-Inc-Korea/AutoRAG

Tue, 28 Jul 2026 05:52:25 +0000

AutoRAG is a MIT librarian agent that searches local document collections with built-in BM25 and MinSync retrieval, optional Jikji find-first discovery, and a CLI that returns curated numbered knowledge units instead of raw grep lines. RepoRadar tier: Gold.

yetone/kill-ai-slop

Tue, 21 Jul 2026 08:18:47 +0000

Kill AI Slop is an Apache-2.0 catalogue and agent skill that detects the visual and copy tics of AI-generated web products and proposes (or applies) the clean fix, with a multilingual field-guide site, a dependency-free standalone scanner, and an installable skill entry on the Vercel Labs `skills` npm registry. RepoRadar tier: Silver.

suboss87/FDEOps

Tue, 21 Jul 2026 08:18:47 +0000

FDEOps is an MIT engagement-memory CLI plus `@fde` agent skill that runs locally as a plain-markdown fieldbook under `~/fde-engagements//.fde/`, so forward-deployed engineers and consultants stop re-explaining the same client context to a new coding-agent session every morning. RepoRadar tier: Silver.

plandex-ai/plandex

Tue, 21 Jul 2026 07:06:06 +0000

Plandex is an MIT AI coding agent built for large codebases and real project work, with a concrete CLI release train, local or self-hosted operating modes, and documentation aimed at editing, planning, and managing bigger engineering tasks than lightweight single-file copilots handle well. RepoRadar tier: Gold.

SYSTRAN/faster-whisper

Tue, 21 Jul 2026 07:06:06 +0000

faster-whisper is an MIT local transcription stack that reimplements Whisper on top of CTranslate2, ships as a maintained Python package, and gives builders a practical path to faster on-device or self-managed speech transcription work. RepoRadar tier: Gold.

microsoft/promptflow

Tue, 21 Jul 2026 07:06:06 +0000

Promptflow is Microsoft's MIT toolkit for building and operating LLM workflows, with a Python package, CLI, docs, and VS Code extension covering prototyping, checking, deployment, and monitoring instead of stopping at notebook demos. RepoRadar tier: Gold.

dyndynjyxa/aio-coding-hub

Tue, 21 Jul 2026 05:11:49 +0000

AIO Coding Hub is an MIT local-first desktop app and localhost gateway that routes Claude Code, Codex, and Gemini CLI traffic through one machine-local control plane, with provider failover, usage telemetry, workspace isolation, plugin hooks, and packaged installs for Windows, macOS, and Linux. RepoRadar tier: Silver.

e2b-dev/open-computer-use

Tue, 21 Jul 2026 03:07:43 +0000

Open Computer Use is an Apache-2.0 stack for running open-source desktop agents inside an E2B cloud Linux sandbox, with live streaming, interchangeable grounding and action models, and explicit keyboard, mouse, and shell control over the sandboxed machine. RepoRadar tier: Gold.

GradientHQ/parallax

Tue, 21 Jul 2026 03:07:43 +0000

Parallax is an Apache-2.0 distributed model-serving framework for building your own inference cluster across mixed nodes, with a documented install path and explicit focus on decentralized serving instead of a single-box local runtime. RepoRadar tier: Gold.

SirAllap/agentglass

Tue, 21 Jul 2026 03:07:43 +0000

agentglass is an MIT mission-control dashboard and workspace for local coding agents that can ingest Claude Code hooks or OpenTelemetry GenAI traces, show live agent sessions, surface spend and tool-call telemetry, and open a diff-review workspace in the same cockpit. RepoRadar tier: Silver.

Orkas-AI/Orkas

Tue, 21 Jul 2026 00:07:40 +0000

Orkas is an MIT local-first desktop client that lets one commander agent coordinate specialist agents, local files, provider-backed model calls, and external coding CLIs like Claude Code, Codex, OpenCode, and Cline from one cross-platform workspace. RepoRadar tier: Silver.

Stack-Cairn/LiveAgent

Mon, 20 Jul 2026 23:19:29 +0000

LiveAgent is an MIT-licensed local-first desktop agent that combines multi-model chat, local tool execution, MCP and Skills support, persistent memory, and an optional browser WebUI for remotely controlling the same runtime. RepoRadar tier: Silver.

agentlas-ai/Agentlas-OS

Mon, 20 Jul 2026 23:19:29 +0000

Agentlas OS is an Apache-2.0 local-first agent hub that creates specialist agents or teams from plain-language requests, runs them across supported model hosts, exposes browser execution surfaces, and can keep owner-scoped agent packages in a private cloud account. RepoRadar tier: Silver.

sebsto/wispr