GitHub Radar

“Sovereign AI” can refer to data location, infrastructure ownership, model licences, organizational control or legal jurisdiction. EuLLM combines several of those ideas in an open-source project aimed at European deployments. Its current engine runs GGUF models locally and exposes familiar APIs, while its broader roadmap promises model verticalization, an EU-hosted hub and compliance-oriented documentation. The project is timely and technically interesting. It must also be evaluated by separating available software from future plans and marketing claims from legal obligations. What EuLLM provides today EuLLM’s current engine is a Rust-based local inference runtime built around llama.cpp. It can run GGUF models and expose Ollama-compatible and OpenAI-compatible APIs. The project publishes binaries for several operating-system and hardware combinations and documents builds with CPU and GPU backends. The repository is Apache-2.0 licensed and describes zero telemetry for the engine. Local ...

An AI coding agent becomes part of a developer’s working environment. It reads source code, proposes patches and may run commands. Locking that workflow to one model provider can make pricing, privacy and model changes harder to manage. OpenCode’s answer is an open-source agent built around provider choice and a terminal-first experience. Provider flexibility is useful, but the most important questions remain what the agent can access and what evidence proves its changes are correct. What OpenCode is OpenCode is an MIT-licensed coding agent with a terminal user interface, language-server integration and a client/server architecture. It can work with several commercial model providers and local models rather than requiring one proprietary backend. Desktop applications are also available in beta. Its language-server support helps the agent work with code structure and diagnostics rather than relying only on raw text. The client/server design allows the agent process and interface to evol...

AI coding tools are often used as isolated conversations: ask for a feature, review the diff and move on. Compound Engineering proposes a different model. It packages repeatable agents and skills so research, planning, implementation, review and accumulated project knowledge can reinforce one another. That structure can make an AI workflow more consistent. It can also add a large instruction and tool surface that teams need to understand. What the plugin provides The Compound Engineering Plugin is an MIT-licensed collection designed for Claude Code, Codex, Cursor and other coding-agent environments. Its documentation lists more than 50 agents and more than 38 skills, covering strategy, planning, code simplification, reviews, documentation, release notes and specialized development patterns. The core idea is compounding knowledge. Instead of treating each task as a blank slate, the workflow records decisions and reusable solutions so later work can start with better context. Setup and d...

Voice agents combine several systems that can fail independently: telephony, speech recognition, language models, text-to-speech, workflow logic and real-time transport. Dograh brings those pieces into an open-source platform with a visual builder and self-hosting option. That control is valuable, particularly for teams that do not want a single voice-AI vendor to own the full stack. It does not make phone automation simple or legally risk-free. What Dograh is Dograh is a platform for building inbound and outbound voice agents. It supports configurable LLM, speech-to-text and text-to-speech providers, telephony integration, visual workflows and self-hosted deployment. Its repository is BSD-2-Clause licensed and the latest verified release was v1.31.0. The documented quick start uses Docker Compose. It also notes that anonymous usage telemetry is enabled in the example command and can be disabled through configuration. That is an important detail for privacy-sensitive deployments. Dogra...

Building an LLM demo is easy compared with operating one. Once an application reaches users, teams need to know which model handled a request, how long it took, what it cost, whether the answer was useful and what changed after a prompt or provider update. TensorZero tries to bring those concerns into one open-source system. Its scope is ambitious: gateway, observability, evaluation, experimentation and optimization. That can reduce fragmentation, but it also creates a platform that must be designed and operated carefully. What TensorZero is TensorZero sits between an application and model providers. Requests pass through a common gateway that can route across supported services and self-hosted inference systems. The same platform records structured inference data, supports evaluations and experiments, and connects production feedback to later prompt or model improvements. The project is provider-agnostic rather than tied to one commercial API. This is useful when teams want fallback r...

Customer relationship management software tends to become infrastructure. Contacts, sales activity, email history, permissions and automation accumulate until moving to another platform becomes a major project. Twenty argues that this category should be more open: a modern CRM with APIs, customization and a path to self-hosting. That makes Twenty more interesting than a visual copy of Salesforce. It also means buyers need to evaluate the entire operational model, not only the interface. What Twenty is Twenty is an open-source CRM built with TypeScript, NestJS, PostgreSQL, Redis and React. It supports familiar CRM work such as contacts, companies, opportunities, views and workflows, while allowing teams to define custom objects and fields. Developers can work through REST and GraphQL APIs, webhooks and an application framework. The project’s positioning increasingly includes AI assistants and agents, but its core value remains a configurable system of record for customer data. Teams can...

Files are often the least convenient part of an AI workflow. A PDF may look clean to a reader but contain broken reading order. A PowerPoint deck can hide important text inside shapes. A spreadsheet mixes values, formulas and structure that do not translate naturally into a prompt. Microsoft’s MarkItDown approaches the problem with a practical target: convert supported documents into Markdown that software and language models can process. That goal makes MarkItDown useful, but it also defines its limits. It is a text-extraction and conversion tool, not a guarantee that every visual detail will survive. What MarkItDown does MarkItDown is a Python package and command-line tool for converting PDFs, Word documents, PowerPoint presentations, spreadsheets and other supported inputs into Markdown. The output can be searched, indexed, reviewed in version control or passed into retrieval and AI pipelines. The basic command is deliberately simple: point the tool at a file and redirect or sav...

CapCut proved that video editing could feel approachable without looking like a simplified toy. It also pushed many creators toward a product whose most useful features, cloud services, and platform decisions are controlled by one company. OpenCut takes the opposite approach: build an editor in public, keep it free and open source, and move as much of the editing workflow as possible onto the user's device. That makes OpenCut one of the most interesting open creator tools on GitHub. It does not, however, make it a complete CapCut replacement today. What OpenCut is OpenCut is an MIT-licensed video editor for the web, with desktop and mobile ambitions. Its current architecture combines a Next.js application with a growing Rust core. GPU compositing, effects, masks, timing, and WebAssembly bindings are being moved into Rust so the same foundation can eventually support browser, desktop, automation, and plugin use cases. The hosted editor is the easiest way to try it. Developers c...

GitHub's trending list moves quickly, but the most useful projects are not always the ones with the biggest one-day star count. For this edition of Weekly Radar, we looked for projects with a clear purpose, an identifiable licence, usable documentation, and enough technical substance to justify attention beyond the trend cycle. The result is a varied group: an ambitious virtual companion, a provider-neutral LLM library, a modular voice-agent stack, a local-first GIS platform, and a WebGPU building editor. All five are open source, but their maturity and operating costs differ significantly. 1. Project AIRI: an open digital companion with a very large roadmap Repository: https://github.com/moeru-ai/airi Licence: MIT Best for: developers experimenting with AI characters, real-time voice, avatars, and game integrations Pricing: open-source software; model, speech, and hosted API providers may charge separately Project AIRI aims to create a self-hosted digital companion that ca...