These Apache or MIT licensed tools cover the lifecycle of AI projects, from knowledge management to agent orchestration. OpenClaw, Superpowers, Hermes, ECC, DeepSeek Harness, the skill Claude Code by Andrej Karpathy... On GitHub, the agent toolkit currently has unparalleled popularity. We expand here the perspective to the entire set of bricks in the AI ecosystem: heterogeneous inference, routing gateways, knowledge management, etc. Here are some of them whose repositories are active. They are classified by number of stars. MarkItDown is a MIT licensed Python utility that converts various file formats into Markdown. Signed by Microsoft, it is comparable to textract, with a focus on preserving the structure (entities, lists, tables, links). Some plug-in are available, including for OCR and for integration with Azure cognitive services. These include, among other things, video and structured field management. OpenHands Agent Canvas is a front-end self-hosted (package npm) under MIT license for managing coding agents compatible with ACP. It uses the OpenHands agent by default, but can use others, on local or remote back-end. One can attach an automation server (GitHub, Slack, Linear...) and an ingress service. Colibri is an inference engine focused on expert mixture models (MoE), developed by Vincenzo Fornaro, a computer vision engineer. Under the Apache 2.0 license, it manages back-ends CPU, CUDA, Metal and NUMA. Its main feature is to assemble RAM, VRAM and NVMe into a single storage hierarchy, with a just-in-time operation distributing the parameters. Main model families supported: DeepSeek, GLM, Inkling, Kimi, OLMoE and Qwen. WeKnora is a knowledge management tool developed by Tencent. Under the MIT license, it combines hybrid search (lexical and semantic; optionally agent-based) and automated creation of wikis with knowledge graphs. Models, vector bases and back-ends of storage are interchangeable. An extension allows to control web browsers. Various integrations exist with productivity tools (Confluence, GitLab, Notion...) and instant messengers. WeKnora has an MCP server, a skill ClawHub and a plug-in for the DeepSeek harness. 9Router is a local LLM gateway under the MIT license. It automatically switches between three billing levels: Claude Code, Codex, GitHub Copilot and Gemini subscriptions; economic models (GLM, Kimi, MiniMax); free versions of Kiro, OpenCode and Vertex AI. You can use multiple accounts per provider, with automatic load balancing. And chain several into a "virtual provider". 9Router manages quota tracking and automatic token OAuth refresh. It can perform local embedding and compress the results of tools before sending them to the LLM. Obscura is a headless browser under the Apache 2.0 license. It is presented as an alternative to Chrome, less memory-intensive, with faster initialization and better session isolation. Obscura is compatible with the Chrome DevTools protocol. You can connect Puppeteer and Plawright to it. Cloudflare cites it as an inspiration in the development of Kitesurf, its own headless browser. Llamafile is based on the APE (Actually Portable Executable) format and aims to make C a truly "portable" language by reconfiguring GCC and Clang to produce a POSIX format usable on multiple OS and CPU architectures. Applied to the framework llama.cpp, this gives Llamafile, which allows the distribution of LLM (weights + inference code) in a single file, executable without installation. Under the Apache 2.0 license, Llamafile exists in a "light" version, without embedded CUDA and Vulkan libraries. To circumvent the 4 Go limit of executable files under Windows, it allows the use of external weights. The latest version of the server manages a catalog of a dozen models, including Mistral and Qwen. WebLLM is another inference engine, which targets browsers via WebAssembly and WebGPU. Under the Apache 2.0 license, it manages Web Workers, allowing tasks to be offloaded to separate threads. This is a project by the MLC (Machine Learning Compilation) community, originally including OctoML and Carnegie-Mellon. It exploits the principle of ML compilation (capturing a tensor operations graph then transformed into low-level representations) to favor heterogeneous execution. ChatGPTBox is a browser extension (Chromium, Firefox, Safari) under the MIT license providing various ways to access ChatGPT: floating discussion windows, dedicated page, actions in the context menu (summary, translation, sentiment analysis, code explanation...). ChatGPTBox has specific integrations with certain sites (GitHub, Reddit, Stack Overflow, YouTube, search engines...). They can be disabled. It can be connected to third-party models in the OpenAPI format and to Ollama servers. Local Deep Research is a local research assistant under the MIT license with several modes of operation: quick summary, detailed search, report creation, document analysis, question-answer. It can use several strategies, including an autonomous one based on LangGraph. Local Deep Research includes a REST API, an MCP server, a dashboard and evaluations. Designed to work on an RTX 3090, it manages DeepSeek, Gemma, Llama, Mistral and Qwen models, via Ollama, LM Studio and llama.cpp. OpenWhispr is a speech recognition application under the MIT license for Windows, Mac and Linux. It can process audio locally (Whisper, Orukeet, NVIDIA Parakeet and Cohere Transcribe models; unlimited use) or online (GPT-5, Claude, Gemini, Groq, Tinfoil, OpenRouter; limited to 2000 words per week). Diarization is done locally. OpenWhispr can automatically transcribe Zoom, Teams and FaceTime meetings, in addition to imported files. You can send screen captures as context. All transcribed elements are stored in a SQLite database with the option of cloud synchronization. By default, the responses appear where the cursor is. They go otherwise to the clipboard. Code2Prompt is a tool under the MIT license developed by Olivier D’Arcona, a Swiss ML engineer and doctoral student at the University of Fribourg (thesis on LLM analysis of documents in finance and law). The tool promises to optimize the conversion of codebases into prompts, allowing customization via templates Handlebars and Glob patterns. It is available in API, DSK Python, skill and MCP server. stable-diffusion.cpp is a pure C/C++ inference approach applied to diffusion models - image and video - based on GGML, for a variety of back-ends (CPU, CUDA, Vulkan, Metal, OpenCL, SYCL). Under the MIT license, stable-diffusion.cpp has various extensions, including: Upscaling with ESRGAN; Editing SDXL images with PhotoMaker; Pattern transfer with IP-Adapter; Object removal with ADetailer (which uses YOLOv8). GodMode is a desktop application (Windows, Mac, Linux) under the MIT license that brings together the web versions of various chatbots (ChatGPT, Claude, Perplexity, Bing, etc.) on a single interface. Its particularity: it sends the prompts to them simultaneously. The web apps have been preferred to APIs because some features are integrated in advance, or are even reserved for them. This also allows covering services that do not have an API (Poe, for example) and those that have one only for certain models (Perplexity). Awesome Prompt Engineering and Awesome Local LLM lists are GitHub repositories that compile links related to prompt engineering/context engineering and local AI. Their content is similar: reference articles, tools, models, datasets, benchmarks, tutorials...