For several decades now, the most innovative software has always emerged from the world of open source software. It’s no different with machine learning and large language models. If anything, the open source ecosystem has grown richer and more complex, because now there are open source models to complement the open source code.
For article, we’ve pulled together some of the most intriguing and useful projects for AI and machine learning. Many of these are foundation projects, nurturing their own niche ecology of open source plugins and extensions. Once you’ve started with the basic project, you can keep adding more parts.
Most of these projects offer demonstration code, so you can start up a running version that already tackles a basic task. Additionally, the companies that build and maintain these projects often sell a service alongside them. In some cases, they’ll deploy the code for you and save you the hassle of keeping it running. In others, they’ll sell custom add-ons and modifications. The code itself is still open, so there’s no vendor lock in. The services simply make it easier to adopt the code by paying someone to help.
Here are 16 open source projects that developers can use to unlock the potential in machine learning and large language models of any size—from small to large, and even extra large.
Agent Skills
AI coding agents are often used to tackle standard tasks like writing React components or reviewing parts of the user interface. If you are writing a coding agent, it makes sense to use vetted solutions that are focused on the task at hand. Agent Skills are pre-coded tools that your AI can deploy as needed. The result is a focused set of vetted operations capable of producing refined, useful code that stays within standard guidelines. License: MIT.
Awesome LLM Apps
If you are looking for good examples of agentic coding, see the Awesome LLM Apps collection. Currently, the project hosts several dozen applications that leverage some combination of RAG databases and LLMs. Some are simple, like a meme generator, while others handle deeper research like the Journalist agent. The most complex examples deploy multi-agent teams to converge upon an answer. Every application comes with working examples for experimentation, so you can learn from what’s been successful in the past. Altogether, the apps in this collection are great inspiration for your own projects. License: Apache 2.0.
Bifrost
If your application requires access to an LLM service, and you don’t have a particular one in mind, check out Bifrost. A fast, unified gateway to more than 15 LLM providers, this OpenAI-compatible API quickly abstracts away the differences between models, including all the major ones. It includes essential features like governance, caching, budget management, load balancing, and it has guardrails to catch problems before they are sent out to service providers, who will just bill you for the time. With dozens of great LLM providers constantly announcing new and better models, why limit yourself? License: Apache 2.0.
Claude Code
If the popularity of AI coding assistants tells us anything, it’s that all developers—and not just the ones building AI apps—appreciate a little help writing and reviewing their code. Claude Code is that pair programmer. Trained on all the major programming languages, Claude Code can help you write code that is better, faster, and cleaner. It digests a codebase and then starts doing your bidding, while also making useful suggestions. Natural language commands plus some vague hand waving are all the Anthropic LLM needs to refactor, document, or even add new features to your existing code. License: Anthropic’s Commercial TOS.
Clawdbot
Many of the tools in this list help developers create code for other people. Clawdbot is the AI assistant for you, the person writing the code. It integrates with your desktop to control built-in tools like the camera and large applications like the browser. A multi-channel inbox accepts your commands through more than a dozen different communication channels including WhatsApp, Telegram, Slack, and Discord. A cron job adds timing. It’s the ultimate assistant for you, the ruler of your data. If AI exists to make our lives easier, why not start by organizing the applications on your desktop? License: MIT.
Dify
For projects that require more than just one call to an LLM, Dify could be the solution you’ve been looking for. Essentially a development environment for building complex agentic workflows, Dify stitches together LLMs, RAG databases, and other sources. It then monitors how they perform under different prompts and parameters and puts it all together in a handy dashboard, so you can iterate on the results. Developing agentic AI requires rapid experimentation, and Dify provides the environment for those experiments. License: Modified version of Apache 2.0 to exclude some commercial uses.
Eigent
The best way to explore the power and limitations of an agentic workflow is to deploy it yourself on your own machine, where it can solve your own problems. Eigent delivers a workforce of specialized agents for handling tasks like writing code, searching the web, and creating documents. You just wave your hands and issue instructions, and Eigent’s LLMs do their best to follow through. Many startups brag about eating their own dogfood. Eigent puts that concept on a platter, making it easy for AI developers to experience directly the abilities and failings of the LLMs they’re building. License: Apache 2.0.
Headroom
Programmers often think like packrats. If the data is good, why not pack in some more? This is a challenge for code that uses an LLM because these services charge by the token, and they also have a limited context window. Headroom tackles this issue with agile compression algorithms that trim away the excess, especially the extra labels and punctuation found in common formats like JSON. A big part of designing working AI applications is cost engineering, and saving tokens means saving money. License: Apache 2.0.
Hugging Face Transformers
When it comes to starting up a brand-new machine learning project, Hugging Face Transformers is one of the best foundations available. Transformers offers a standard format for defining how the model interacts with the world, which makes it easy to drop a new model into your working infrastructure for training or deployment. This means your model will interact nicely with all the already available tools and infrastructure, whether for text, vision, audio, video, or all of the above. Fitting into a standard paradigm makes it much easier to leverage your existing tools while focusing on the cutting edge of your research. License: Apache 2.0.
LangChain
For agentic AI solutions that require endless iteration, LangChain is a way to organize the effort. It harnesses the work of a large collection of models and makes it easier for humans to inspect and curate the answers. When the task requires deeper thinking and planning, LangChain makes it easy to work with agents that can leverage multiple models to converge upon a solution. LangChain’s architecture includes a framework (LangGraph) for organizing easily customizable workflows with long-term memory, and a tool (LangSmith) for evaluating and improving performance. Its Deep Agents library provides teams of sub-agents, which organize problems into subsets then plan and work toward solutions. It is a proven, flexible test bed for agentic experimentation and production deployment. License: MIT.
LlamaIndex
Many of the early applications for LLMs are sorting through large collections of semi-structured data and providing users with useful answers to their questions. One of the fastest ways to customize a standard LLM with private data is to use LlamaIndex to ingest and index the data. This off-the-shelf tool provides data connectors that you can use to unpack and organize a large collection of documents, tables, and other data, often with just a few lines of code. The layers underneath can be tweaked or extended as the job requires, and LlamaIndex works with many of the data formats common in enterprises. License: MIT.
Ollama
For anyone experimenting with LLMs on their laptop, Ollama is one of the simplest ways to download one or more of them and get started. Once it’s installed, your command line becomes a small version of the classic ChatGPT interface, but with the ability to pull a huge collection of models from a growing library of open source options. Just enter: ollama run and the model is ready to go. Some developers are using it as a back-end server for LLM results. The tool provides a stable, trustworthy interface to LLMs, something that once required quite a bit of engineering and fussing. The server simplifies all this work so you can tackle higher level chores with many of the most popular open source LLMs at your fingertips. License: MIT.
OpenWebUI
One of the fastest ways to put up a website with a chat interface and a dedicated RAG database is to spin up an instance of OpenWebUI. This project knits together a feature-rich front end with an open back end, so that starting up a customizable chat interface only requires pulling a few Docker containers. The project, though, is just a beginning, because it offers the opportunity to add plugins and extensions to enhance the data at each stage. Practically every part of the chain from prompt to answer can be tweaked, replaced, or improved. While some teams might be happy to set it up and be done, the advantages come from adding your own code. The project isn’t just open source itself, but a constellation of hundreds of little bits of contributed code and ancillary projects that can be very helpful. Being able to customize the pipeline and leverage the MCP protocol supports the delivery of precision solutions. License: Modified BSD designed to restrict removing OpenWebUI branding without an enterprise license.
Sim
The drag-and-drop canvas for Sim is meant to make it easier to experiment with agentic workflows. The tool handles the details of interacting with the various LLMs and vector databases; you just decide how to fit them together. Interfaces like Sim make the agentic experience accessible to everyone on your team, even those who don’t know how to write code. License: Apache 2.0.
Sloth
One of the most straightforward ways to leverage the power of foundational LLMs is to start with an open source model and fine-tune it with your own data. Unsloth does this, often faster than other solutions do. Most major open source models can be transformed with reinforcement learning. Unsloth is designed to work with most of the standard precisions and some of the largest context windows. The best answers won’t always come directly from RAG databases. Sometimes, adjusting the models is the best solution. License: Apache 2.0.
vLLM
One of the best ways to turn an LLM into a useful service for the rest of your code is to start it up with vLLM. The tool loads many of the available open source models from repositories like Hugging Face and then orchestrates the data flows so they keep running. That means batching the incoming prompts and managing the pipelines so the model will be a continual source of fast answers. It supports not just the CUDA architecture but also AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, Arm CPUs, and TPUs. It’s one thing to experiment with lots of models on a laptop. It’s something else entirely to deploy the model in a production environment. vLLM handles many of the endless chores that deliver better performance. License: Apache-2.0.
