17 Cutting‑Edge Open‑Source Projects From AI Agents to NanoGPT

Most of the tools listed either wrap an MCP server, communicate with it, or assume one is already running, turning the author’s MCP Shelf Directory into core infrastructure rather than a niche index.

AgentClick is an HTTP interception layer placed between AI agents such as Claude Code and their execution environment, enforcing a manual approval step before any operation. It provides a browser UI for reviewing, editing, or approving proposals involving email drafts, code diffs, or multi‑step plans, persists review feedback per session, and learns preferences over time.

View on GitHub: https://github.com/agentlayer-io/AgentClick

ts‑graph‑mcp solves the problem that AI agents lack a structured way to understand symbol relationships, call graphs, or inheritance chains in large TypeScript codebases. By performing AST analysis, it parses the codebase into a queryable graph database, extracts functions, classes, and interfaces, and adds semantic search. On first run it downloads an embedding model to index each symbol by concept, exposing the graph via the searchGraph MCP tool for Claude Code to query by meaning rather than exact name.

View on GitHub: https://github.com/couzic/ts-graph-mcp

ContextEngine is an MCP server that addresses the painful issue of AI programming agents forgetting everything between sessions and repeatedly making the same architectural mistakes. Running entirely locally without API keys, it uses the all‑MiniLM‑L6‑v2 model for mixed keyword‑and‑semantic search, extracts relevant rules from files such as CLAUDE.md and copilot‑instructions.md, and enforces protocol discipline via a pre‑commit hook.

View on GitHub: https://github.com/FASTPROD/ContextEngine

nanogpt is an ultra‑minimal GPT‑2 re‑implementation that reproduces the 124M‑parameter model on a single 8×A100 node in about four days using the OpenWebText dataset. The repository contains only two files— train.py and model.py —each under 300 lines, offering a pure PyTorch training loop, optional OpenAI weight loading, and plug‑and‑play support for custom datasets via Hugging Face and NumPy.

View on GitHub: https://github.com/ttwthomas/nanogpt

claude‑code is a TypeScript/Bun CLI tool that puts Claude’s language model directly in the terminal, eliminating context switches between editor, browser, and chat UI. It reads the actual codebase, runs git workflows, and handles everyday programming tasks through simple natural‑language commands. The architecture comprises roughly 1,900 source files organized around a modular tool system, command layer, and service modules, with clear permission controls and feature flags.

View on GitHub: https://github.com/codeaashu/claude-code

Scheduled is an open‑source AI scheduling agent that removes the back‑and‑forth of email‑based meeting coordination. It reads incoming threads, checks calendar availability, and drafts reply suggestions with suggested times, handling time‑zones, buffers, and group meetings automatically. Drafts are never sent without explicit user approval, and the system runs self‑hosted using the user’s own API keys.

View on GitHub: https://github.com/Fergana-Labs/scheduler

nanoGPT (second entry) reiterates the same minimal‑code philosophy: two ~300‑line Python files that train a GPT‑2‑style transformer on a single 8×A100 node in roughly four days. Because the code is readable, modifying architecture, swapping custom data, or experimenting with training schedules involves editing plain Python rather than digging through deep framework abstractions.

View on GitHub: https://github.com/flexaihq/nanoGPT

DrunkenViking is a grid‑puzzle game built for Mini Ludum Dare 53 that demonstrates a clean solution to rapid prototype constraints: implementing turn‑based movement, collision detection, and physics‑influenced character behavior without accruing architectural shortcuts that become technical debt. The source remains tidy and serves as a reference for developers needing a lightweight tile‑map level‑scaffolding.

View on GitHub: https://github.com/cxong/DrunkenViking

LocalAI runs LLM, vision, speech, and image models entirely on local hardware, removing hard dependencies on cloud AI APIs and even GPUs. It natively supports OpenAI and Anthropic API dialects, 36+ backends (including llama.cpp and transformers), runs on NVIDIA, AMD, Intel, Apple Silicon, and CPU‑only systems, and includes built‑in RAG, tool usage, and multi‑user role‑based access control, ensuring data never leaves the machine.

View on GitHub: https://github.com/mudler/LocalAI

open‑multi‑agent is a TypeScript orchestration engine that eliminates the manual graph‑wiring required by tools like LangGraph. By calling runTeam() with a target, it automatically handles task decomposition, dependency resolution, and parallel execution. Running on Node.js with only three runtime dependencies, it can compose hybrid model teams across Claude, GPT, Gemini, and local models in a single run, and can be dropped into existing back‑ends without rebuilding the agent stack.

View on GitHub: https://github.com/JackChen-me/open-multi-agent

claude‑agent‑sdk‑python solves the problem of integrating Claude Code into Python workflows without managing subprocesses or CLI pipelines. It exposes two concise asynchronous interfaces—a single‑prompt query function and a stateful multi‑turn conversation client with custom tool hooks. Fine‑grained permission, whitelist/blacklist, and working‑directory configuration are available per call, and the SDK installs via pip for Python 3.10+.

View on GitHub: https://github.com/anthropics/claude-code-sdk-python

vikingcode.github.io is a static portfolio and documentation site that strips away heavyweight build pipelines and component libraries, relying only on plain HTML, CSS, and JavaScript with zero unnecessary dependencies. It offers a clean baseline that can be forked or used for extracting patterns, leveraging semantic markup and true browser compatibility for technical documentation.

View on GitHub: https://github.com/vikingcode/vikingcode.github.io

anthropic‑sdk‑go is the official Go client for Anthropic’s Claude API, removing the need to hand‑write HTTP clients, auth headers, and stream parsers. It handles request serialization, API authentication, and streaming responses in idiomatic Go, with an active contributor base of 18 and a single go get install command.

View on GitHub: https://github.com/anthropics/anthropic-sdk-go

openharness is a local‑terminal CLI that places an AI programming agent directly in the command line, avoiding vendor lock‑in and supporting any LLM, local model, or cloud API interchangeably. It addresses the specific issue of most agent tools binding users to a single vendor and lacking recovery paths when automation fails. Features include checkpoints, rollback, permission‑based execution, 42 integrations, and 890 tests ensuring strict TypeScript correctness.

View on GitHub: https://github.com/zhijiewong/openharness

weekly tackles the fragmentation of technical signals by curating a weekly GitHub‑based newsletter that aggregates AI breakthroughs, software trends, and industry analysis. With 467 contributors submitting resources via issues and pull requests, it also runs a free developer job board where engineers can post openings without fees or barriers.

View on GitHub: https://github.com/ruanyf/weekly

claude‑code‑source publishes the internal mechanisms of Claude Code v2.1.88, a terminal‑based agent programming tool that handles git operations, code‑base analysis, and task execution via natural‑language commands. The repository addresses the core problem of opaque source‑mapping and navigation during active debugging sessions, offering improved source‑mapping capabilities for developers encountering non‑transparent behavior across projects.

View on GitHub: https://github.com/hangsman/claude-code-source

The overarching observation is that many of these projects form the scaffolding for AI agents rather than direct developer tools. AgentClick, ContextEngine, ts‑graph‑mcp, and open‑multi‑agent all address the same upstream challenges: agents lacking sufficient visibility, memory, or trust to operate unsupervised. From a regulatory perspective this space is still nascent. While AgentClick’s browser‑based approval layer and OpenHarness’s local CLI represent opposite ends of the stack, no single approach has yet demonstrated clear superiority. nanoGPT stands out as an outlier, offering a clean, self‑contained training loop without any MCP wrapper, making it the most transparent example of an open‑source AI project in the list.