Explore the directory

An OpenCode plugin that runs a local HTTP gateway translating between the OpenAI, Anthropic and Gemini API formats and whatever LLM provider OpenCode has configured, including streaming and parallel tool calling, so other tools only need one local endpoint.

• LLM gateways

An open-source pipeline for training natural language understanding models for text classification and named entity recognition, which automatically chooses between few-shot and full fine-tuning methods based on how much labeled data is available.

• Models

A scenario-based knowledge base question-answering system that turns uploaded documents into an LLM-generated wiki with a linked knowledge graph, then publishes each scenario as its own public question-answering page or an API-authenticated knowledge base for an AI agent.

• Agent memory
• RAG

A fine-tuning toolkit for adapting general-purpose language models to specialized domains, such as legal, medical or financial text, that analyzes a dataset's token distribution and class balance to propose a fine-tuning strategy while still allowing full manual control.

• Models

A multi-task learning framework built on the modernBERT architecture that trains several NLP tasks together in one shared network, with dynamic loss-weighting and task scheduling intended to reduce GPU hours compared with fine-tuning a separate model per task.

• Models

A callback library for Hugging Face's Trainer that forwards training events, such as loss and epoch boundaries, to an observability backend through Pydantic Logfire, without a training script needing to be modified or a custom callback subclass written for each project.

• Models
• Monitoring

A local MCP coding workspace for AI agents accessed through ChatGPT's web interface, letting an agent read files, edit code, run checks and inspect git on the user's own machine through an OpenAI Secure Tunnel, with a tray app and metrics dashboard.

• Coding agents
• MCP servers

An independent research branch of llama.cpp implementing DualDeadline, a component-staged expert-prefetching technique for offloaded mixture-of-experts inference, letting a GPU claim only router-selected expert tensor slices while routing and outputs stay exact.

• Inference
• Papers

A Python framework for building chatbots and AI agent workflows across multiple model providers, including OpenAI, LLaMA, DeepSeek, Mistral and Stable Diffusion, through one unified access layer that also supports the Model Context Protocol.

• Agent frameworks

A progressive, curriculum-style collection of scripts and notebooks for fine-tuning Hugging Face Transformers models, moving from the high-level Trainer API down to a hand-written training loop, tokenizer details, and methods such as LoRA and prefix tuning.

• Models
• Tutorials

An open-source framework from a university research group that wraps existing coding agents, such as Codex or OpenCode, into an evolutionary loop that co-evolves a population of candidate solutions alongside a population of agent guidance and skills.

• Coding agents
• Agent frameworks

A research runtime, written in Rust, that runs mixture-of-experts language models entirely on an Apple Silicon CPU, with hand-written NEON and SME kernels and no GPU code path, benchmarked against llama.cpp with corrected, re-measured comparisons.

• Inference

A Python SDK and set of tools that add runtime governance to AI agents: cost limits, step limits, loop detection and policy enforcement, plus an MCP wrapper and Claude Code hooks, working offline with local YAML policies.

• AI evals

A source-linked index of published AI model benchmark scores that keeps each score tied to its model release, benchmark variant and evaluation setting, without combining results from different harnesses into an invented composite score.

• AI evals

A peer-to-peer network where providers offer AI inference services and buyers discover them through a DHT and route requests through encrypted connections, with a router that scores providers on price, latency and reputation.

• LLM gateways
• Networking

An open-source Go SDK for building AI agents whose state, tool calls and execution loops survive process crashes, running in-process by default or, for distributed execution, through Temporal or Restate.

• Agent frameworks

An open-source coding agent CLI, daemon and Telegram bridge that builds a feature and verifies its own work in the background, notifying the user when it finishes or needs a decision, and works with any model provider including local models.

• Coding agents

A Python multi-agent platform built on a microservice, event-driven architecture, with a hybrid memory combining SQL, ChromaDB and a KuzuDB-based graph store, and Docker-in-Docker sandboxing that runs agent-written code in disposable, isolated containers.

• Agent frameworks
• Agent memory

A Go dependency-injection library that generates plain Go code at compile time, combining a minimal Fx-style API with Wire-style code generation so there is no runtime reflection or extra runtime dependency.

• Libraries

A local, self-hosted platform that generates complete applications, including a web UI, API, database and CLI, from a natural-language description, using more than 30 integrated local services running on the user's own hardware.

• Coding agents
• Agent frameworks

A spec-driven task-management tool for AI coding assistants such as Claude Code, Codex and Gemini CLI, built around reviewing an agent's plan before it writes code rather than reviewing the resulting diff afterward.

• Coding agents

A self-hosted, MCP-based memory server for AI coding agents such as Claude Code, Cursor and Windsurf, that automatically captures, categorizes and retrieves project-scoped decisions, debug sessions and patterns across sessions and restarts.

• MCP servers
• Agent memory

A white-box, self-hosted platform for learning and inspecting retrieval-augmented generation, showing the whole pipeline: document ingestion, chunking, embedding, vector and BM25 retrieval, prompt assembly and citation tracing, backed by its own benchmark data.

• RAG
• Tutorials

A narrow, self-contained inference engine for two specific Qwen models, Qwen3.6-35B-A3B and Qwen3.8-27B, with Vulkan and Metal graph executors tuned for the AMD BC-250 and Apple Silicon, plus a built-in coding agent and HTTP server.

• Coding agents
• Inference