Find the open-source solution that fits.

Project overview

Shadow signs each credit decision with Ed25519 and chains attestations with SHA-256 so a later verifier can detect silent changes to the verdict, reason codes, model manifest, or committed inputs and outputs. The verdict engine uses deterministic rules, while LLM personas provide advisory rationale that cannot alter the decision. The project sends no telemetry and explicitly remains pre-1.0 rather than a compliance product. It is a research layer for lenders building an auditable decision narrative.

Project overview

Browser Use gives an AI agent direct browser interaction for navigation, clicking, typing, and form completion. The CLI is designed for one-off browser tasks driven by an existing agent such as Claude Code or Codex, while the Python library supports repeatable automation, parallel jobs, custom tools, and product integrations with a model provider of the developer's choice.

Project overview

Elixir Desktop lets Phoenix LiveView developers build native-like local applications for Windows, macOS, Linux, iOS, and Android. It routes window, webview, menu, and notification operations through platform backends, using wxWidgets on desktop, a JSON bridge on mobile, or a browser backend for headless environments. An example application and setup guides provide a starting point. Core desktop and mobile execution works, but complete desktop installer generation, signing, and one-command cross-platform packaging remain roadmap items.

Project overview

smolagents offers a compact Python framework for code-writing agents, conventional tool-calling agents, and multi-agent hierarchies, with support for different model providers and Hub sharing. Its CodeAgent expresses actions as Python snippets. Since those snippets are executed, untrusted workloads should use a sandboxed backend; the local Python executor is explicitly not a security boundary.

Project overview

Agent Lightning connects existing agents to optimization workflows with minimal code changes by collecting prompts, tool calls, and rewards as structured traces. It can selectively optimize one or more agents built with LangChain, the OpenAI Agent SDK, AutoGen, CrewAI, Microsoft Agent Framework, or plain Python. Supported approaches include reinforcement learning, automatic prompt optimization, and supervised fine-tuning.

Project overview

Install and update Agent Skills from GitHub, local folders, or ZIP files, validate them against the specification, and manage their metadata from one toolkit. SkillPort can deliver skills through a CLI or MCP server and uses search-first, on-demand loading for large collections. It is still under active development, so its APIs may change.

Project overview

SWE-smith creates Docker execution environments from GitHub repositories, synthesizes tasks such as file localization and program repair, filters instances that break one or more tests, and generates issue text. The resulting tasks and trajectories can support SWE-agent training or evaluation. It requires Docker and was developed for Ubuntu 22.04; Windows and macOS are not planned targets.

Project overview

React NodeGUI is a React renderer for NodeGUI that builds Qt5-based desktop applications for Windows, macOS, and Linux. Developers can compose native widgets with React, style them with cascading CSS and Flexbox, and use Node.js APIs without bundling a Chromium runtime. The repository explicitly states that the project is no longer maintained, so adopting it now means accepting an aging stack or maintaining a fork yourself.

Project overview

Mirobody has a C++ core that runs inside Android and iOS apps, in a desktop wrapper, or as a self-hosted server. It brings wearable, laboratory, and clinical records behind a common interface, with FHIR endpoints, long-term memory, MCP tools, and several LLM providers. Health data can remain on the phone or the operator’s own infrastructure, but enabling an external model, synchronization, or vendor integration can send selected information outside that boundary. The architecture supplies privacy and compliance controls; the deployer remains responsible for the final HIPAA or GDPR posture.

Project overview

smol developer turns a product specification into a shared plan, a file list, and generated code for each file. It can run as a command-line scaffold, be imported as a Python library, or be exposed through an Agent Protocol API inside another application. The workflow is intentionally human-in-the-loop, and the project documents unresolved limits around cross-file coherence, dependency setup, self-healing, and slow generation.

Project overview

Particle Device OS contains the firmware libraries and system components used to build software for Particle hardware. Developers can use its build guides, application firmware API, debugging support, bootloader, cloud communication modules, threading code, and hardware abstraction layers. It is primarily a firmware development source tree, and changes to LGPLv3-licensed system components carry the sharing obligations described by the project.

Project overview

LibVLCSharp exposes VideoLAN's LibVLC engine to .NET and Mono applications across desktop, mobile, server, and Unity environments. Developers can render video, output audio, handle files and network streams, and use encoding or streaming capabilities through a common API, with integrations for frameworks such as WPF, WinForms, MAUI, Avalonia, and Xamarin. Each target also needs the platform-specific native LibVLC package, because the bindings do not include the media engine by themselves.

Project overview

agent-lsp orchestrates existing language servers and exposes their code intelligence to AI agents through MCP. It keeps language servers warm between sessions, combines many routine calls into batch operations, and supports workflows such as caller analysis and speculative edits. The project documents 65 tools, 24 agent workflows, and CI-verified support for 30 languages.

Project overview

council-diff runs five differently biased personas in parallel on the same decision and exposes their scores, verdicts, strengths, gaps, agreement, and recommendation. Built-in domains cover founder, engineering, investing, career, product, and quant questions, while custom rosters are supported. An optional Oracle reviews difficult or split calls, and the Brier module scores predictions after outcomes resolve. Safe mode can keep the Oracle on a zero-retention model when retention matters.

Project overview

MCP-Zero is the open-source code accompanying a paper on active tool discovery and retrieval for autonomous LLM agents. It includes instructions for an MCP-tools dataset covering 308 servers and 2,797 tools, plus code under MCP-tools/build_data for constructing a custom server dataset. The data file must be obtained separately and placed at the path expected by the evaluation code.

Project overview

AgentFigureGallery gives coding agents a real visual reference library instead of relying on prompt-only descriptions of scientific style. Search the browser gallery, mark examples as liked, rejected, or selected, and export those choices as a bundle for Claude Code, Codex, Cursor, or another agent to use while writing plotting code. An optional full-public pack contains 16,341 candidates across ten plot types, while a small built-in pack supports quick smoke tests. Included example workflows can export PNG, PDF, SVG, and combined previews.

Project overview

sci-figure renders molecules, crystals, slabs, and nanoparticles from common structure data, then composes them with native draw.io elements. Its Python modules cover atom rendering, reusable figure layout, low-level XML building, and export to PNG, SVG, or PDF while keeping the editable source beside each image.

Project overview

ai-security-guard combines zero-dependency TypeScript runtime guards with a Claude skill for prompt injection, replayed-history attacks, privilege escalation, forged action markers, cross-tenant PII leaks, media abuse, and unsafe rendering. It can log blocked attacks and propose new patterns, but each candidate must pass a regression check before a human promotes it, so nothing becomes live automatically. The reference code uses Express-style functions that can be ported to other stacks.