Overview

This week's trending set is dominated by the race to build better AI coding agents and the infrastructure that powers them. Open-source agent toolkits (Pi, Kimi Code, Qwen Code) are competing alongside new agent orchestration layers (Ouroboros, Orca) and MCP-native utilities (fastmcp, code-review-graph, wigolo). Several new entries focus on agent skills and structured output — clear signs the ecosystem is moving from demo agents to production-grade, composable workflows.

1. PrefectHQ/fastmcp

🚀 The fast, Pythonic way to build MCP servers and clients. FastMCP gives you everything you need to go from prototype to production with a simple decorator-based API. Built by Prefect, it's already seeing rapid adoption in the community.

Why it's on the list: Sets the standard for Python MCP development — minimal boilerplate, production-ready.

2. QwenLM/qwen-code

An open-source AI coding agent that lives in your terminal. Comes with Auto-Memory, Auto-Skills, SubAgents, Agent Teams, and MCP support — dynamic workflows with zero setup. The framework and underlying Qwen models are fully open-source.

Why it's on the list: A complete, open-source coding agent stack from a major model provider; agent feature set rivals closed-source alternatives.

3. diegosouzapw/OmniRoute

Free MIT AI gateway: a single endpoint for 290+ providers (90+ free), 500+ models. Works with Claude Code, Codex, Cursor, OpenCode, Cline & Copilot. Features quota-aware auto-fallback, RTK+Caveman token compression (15-95% savings), and MCP/A2A support. Built by 500+ contributors.

Why it's on the list: The most practical AI gateway for agent developers — reduces cost and complexity of multi-model workflows.

4. tirth8205/code-review-graph

Local-first code intelligence graph for MCP and CLI. Builds a persistent, token-aware map of your codebase so AI coding tools read only what matters — benchmarked context reductions on reviews and large-repo workflows.

Why it's on the list: Solves the "context window is too small" problem for real-world codebases; MCP-native design is ahead of the curve.

5. earendil-works/pi

AI agent toolkit: unified LLM API, agent loop, TUI, and a self-extensible coding agent CLI. Includes a session system, skill management, and both a CLI and web UI (via agegr/pi-web). Designed for multi-session, customizable workflows.

Why it's on the list: A full-featured agent harness that bridges the gap between tool and framework — already has a growing ecosystem.

6. MoonshotAI/kimi-code

Kimi Code CLI is the next-generation terminal AI agent from Moonshot AI. It can read and edit code