GitHub Weekly: MCP Tools Trending Right Now (September 28 – October 4, 2026)

TL;DR: This week’s angle is MCP (Model Context Protocol) — the standard that lets AI tools like Claude and ChatGPT plug into outside data and services. The repos below split into two camps: tools for wiring ChatGPT or Codex into other agent setups, and MCP servers that give an agent real, narrow superpowers in one specific domain like SEO or game modding. If you’re building anything agent-adjacent, there’s a mix of plumbing and niche tools worth knowing about here.

Instead of the usual overall trending list, this week we’re spotlighting repos tagged “mcp” that have shown up in the last 90 days — so you’re seeing newer MCP-related projects rather than the same handful of established frameworks every time. Here’s what stood out for the week of September 28 through October 4, 2026.

#1. miuuyy/codex-chatgpt-web

github.com/miuuyy/codex-chatgpt-web — TypeScript, 13,449 stars

Codex is OpenAI’s coding agent, and it normally burns through its own separate API quota every time it runs. This project lets you swap in your existing ChatGPT Web subscription (including ChatGPT Pro) as the model behind Codex instead, while still keeping context, tool calls, streaming responses, and image support working as expected.

Good for: Developers who already pay for ChatGPT Pro and want to use Codex’s coding-agent workflow without paying for a second, separate API quota.

#2. XiaoDuoYa/codex-with-chatgpt

github.com/XiaoDuoYa/codex-with-chatgpt — TypeScript, 7,028 stars

A close cousin of #1 above, with a different split of labor: ChatGPT handles the high-level planning and reasoning, while Codex’s own harness (the part that actually edits files, runs commands, and calls tools) does the execution. It’s a bet that separating “thinking” from “doing” across two models produces better results than asking one model to do both.

Good for: Developers experimenting with different coding-agent architectures who want to compare a planner-plus-executor setup against a single all-in-one agent.

#3. CopilotKit/OpenBot

github.com/CopilotKit/OpenBot — TypeScript, 6,043 stars

OpenBot gives each AI agent its own sandboxed workstation — a browser, a file system, and a set of tools — and logs every action the agent takes before and after it happens, so there’s an audit trail instead of a black box. It works with any agent built on AG-UI, a protocol for connecting agents to user interfaces.

Good for: Teams running autonomous agents in production who need a clear record of what an agent did and why, for debugging or compliance purposes.

#4. KKKKhazix/AIHOT

github.com/KKKKhazix/AIHOT — TypeScript, 5,846 stars

AIHOT is a self-hosted framework for building your own “what’s trending” site. You plug in your own sources and curation rules, and it compiles an automated daily digest — the same basic idea as a Hacker News front page, but scoped to whatever niche or industry you choose instead of general tech news.

Good for: Content teams, newsletter writers, or community builders who want an automated trending-topics feed for their own niche without relying on a third-party aggregator.

#5. totec448-spec/chat-on-steroids

github.com/totec448-spec/chat-on-steroids — TypeScript, 4,239 stars

This project bolts MCP tool-calling onto the regular ChatGPT client through a local bridge, adds Chrome browser integration, and layers on goal tracking plus the ability to pause and resume long multi-step agent workflows without losing state.

Good for: People who want to extend plain ChatGPT with real tool access and durable multi-step workflows instead of switching to a dedicated agent framework.

#6. superdesigndev/treg

github.com/superdesigndev/treg — Python, 4,153 stars

OpenRouter is a well-known service that lets you route a single API call across many different LLM providers. Treg applies that same routing idea to credentials: it’s a registry and proxy that manages API keys and secrets for agent tools, so you’re not hardcoding separate keys into every agent you build.

Good for: Teams running several AI agents that need shared, centrally managed access to tool credentials instead of scattered config files full of secrets.

#7. zvec-ai/zvec-grep

github.com/zvec-ai/zvec-grep — Rust, 3,920 stars

A fast local search tool in the spirit of ripgrep, but built to serve both humans and AI agents. It combines traditional keyword search (BM25, a standard text-ranking algorithm) with semantic vector search, all running locally so your codebase never has to be uploaded to a hosted indexing service.

Good for: Developers and coding-agent builders who want fast, private code search without depending on a cloud service to index their repository.

#8. Ryze-AI-Adgent/open-seo-mcp-skills

github.com/Ryze-AI-Adgent/open-seo-mcp-skills — Shell, 3,806 stars

An MCP server that connects Claude directly to your real Google Search Console, Google Analytics, and ad account data, giving it tools for keyword research, rank tracking, site audits, and backlink analysis. It also tracks “AI visibility” (also called GEO, or generative-engine optimization) — essentially, how often your content shows up when people ask AI chatbots questions.

Good for: Marketing teams or solo site owners who want Claude to run SEO analysis against their own live traffic and search data instead of generic, one-size-fits-all advice.

#9. rehan-remade/universal-modder

github.com/rehan-remade/universal-modder — Python, 3,364 stars

A set of tools and an MCP server that lets Claude Code dig into a PC game’s files, reverse-engineer its assets, generate new art, 3D models, or audio using the fal AI platform, and then test the resulting mod in-game.

Good for: Hobbyist game modders who want an AI assistant to take on the tedious reverse-engineering and asset-creation work that modding usually requires.

#10. fuxicodex/Fuxi

github.com/fuxicodex/Fuxi — Python, 3,350 stars

FuXi is a terminal-based coding agent, similar in spirit to Codex or Claude Code, that can edit code, run commands, and drive tools. What sets it apart is cost-aware routing across LLM providers — it can send simpler requests to cheaper models and save the expensive ones for tasks that actually need them.

Good for: Developers who want a terminal coding agent without locking into a single LLM vendor, especially if keeping API costs down matters.

Put together, this week’s MCP list breaks down into two clear groups: plumbing that lets ChatGPT and Codex talk to each other or to other agent frameworks, and narrow MCP servers that hand an agent real capability in one specific domain, whether that’s SEO data, local code search, or game asset creation. Worth keeping an eye on which of these niche MCP servers keep growing versus which fade once the novelty wears off.