Here’s today’s look at what developers are starring on GitHub, plus the latest product updates from the GitHub changelog. Star counts below are exactly as shown on GitHub Trending and its daily language pages (Python, JavaScript, TypeScript, PHP) on the evening of 15 November 2026 (IST). They change every hour, so expect different numbers when you click through.

The big theme: giving agents a proper interface

Agent tooling still dominates Trending, but today’s picks focus on the plumbing around agents: a protocol for connecting agents to front ends, a React chat UI kit, the fastest way to build MCP servers in Python, and design skills for coding agents. Alongside them are a governed text-to-SQL layer, two self-hosted email projects and the tensor library behind a lot of local AI.

Trending repositories

1. ag-ui-protocol/ag-ui

AG-UI is “the Agent-User Interaction Protocol”, built to bring agents into front-end applications. It defines how an agent backend and a user interface talk to each other, so you don’t have to invent your own event format for streaming, tool calls and state.

  • Language: TypeScript
  • Stars: 17,649 (393 today)
  • Why devs care: every team building an agent product ends up writing custom glue between the agent and the UI. A shared, MIT-licensed protocol makes it easier to swap agent frameworks or front ends without rewriting that layer.

2. assistant-ui/assistant-ui

A TypeScript/React library for AI chat interfaces. It gives you the building blocks of a chat UI (threads, messages, composers) so you can focus on the model and tools behind it.

  • Language: TypeScript
  • Stars: 13,230 (92 today)
  • Why devs care: a polished chat experience with streaming, editing and attachments is a lot of front-end work. Pairing a ready-made React UI kit with a protocol like AG-UI is a quick path to a production-looking assistant.

3. PrefectHQ/fastmcp

“The fast, Pythonic way to build MCP servers and clients.” FastMCP lets you expose Python functions as Model Context Protocol tools and resources with very little boilerplate, and also connect to MCP servers as a client.

  • Language: Python
  • Stars: 28,503 (76 today)
  • Why devs care: MCP has become the standard way to give AI assistants access to your own systems. If you write Python, FastMCP is one of the shortest routes from “I have a function” to “my agent can call it”. See the docs at gofastmcp.com.

4. google-labs-code/stitch-skills

A library of Agent Skills designed to work with the Stitch MCP server. Each skill follows the Agent Skills open standard, so it works with coding agents such as Antigravity, Gemini CLI, Claude Code and Cursor.

  • Language: TypeScript
  • Stars: 9,117 (248 today)
  • Why devs care: it’s a practical, Apache-2.0-licensed example of packaging design know-how as reusable skills that several agents can share, a pattern worth copying for your own team’s conventions.

5. Canner/WrenAI

WrenAI describes itself as GenBI (generative BI) for AI agents: an open-source, governed text-to-SQL system that uses an open context layer to turn natural-language questions into dashboards, charts and SQL across 20+ data sources, including BigQuery, Snowflake, PostgreSQL, ClickHouse, Amazon Redshift and Databricks.

  • Language: Python
  • Stars: 18,261 (52 today)
  • Why devs care: “chat with your data” is easy to demo and hard to trust. The focus on a governed context layer, where you define what tables and metrics mean, is the part that makes text-to-SQL usable for real reporting.

6. maillab/cloud-mail

An all-in-one email service platform powered by Cloudflare, built around Cloudflare Email and Workers.

  • Language: JavaScript
  • Stars: 16,156 (412 today)
  • Why devs care: if you already run domains on Cloudflare, this is a way to host your own mailboxes on serverless infrastructure instead of paying per seat. It was the top gainer on today’s JavaScript page. Read the docs carefully before trusting it with important mail.

7. karatektus/pl_mail

A self-hosted, Gmail-style mail client with multi-account IMAP, threaded conversations and real-time push via IMAP IDLE and Mercure. It’s built with Symfony, FrankenPHP and Hotwire Turbo, and the README promises you can clone it and run docker compose up.

  • Language: PHP
  • Stars: 637 (25 today)
  • Why devs care: beyond being a useful app, it’s a neat reference project for modern PHP: Symfony on FrankenPHP with real-time updates through Mercure and Turbo, with no heavy JavaScript framework. It led today’s PHP trending page by new stars.

8. ggml-org/ggml

The tensor library for machine learning that sits underneath llama.cpp and many other local inference tools.

  • Language: C++
  • Stars: 15,822 (154 today)
  • Why devs care: ggml and llama.cpp recently reached v1.0.0 with semantic versioning, so the low-level library now has a stable versioning story. If you build bindings or custom inference code, this is the layer to watch.

Today on the GitHub changelog

There are no new changelog entries dated 14 or 15 November (it’s the weekend), so here are recent updates not covered in our earlier roundups.

Manage AI Scan for repositories with the REST API (13 Nov)

You can now turn code scanning’s AI Scan for repositories on or off through the REST API at organization and repository level. This public preview adds a repository_scan setting to the existing endpoints, GET/PATCH /orgs/{org}/code-scanning/ai-scan and GET/PATCH /repos/{owner}/{repo}/code-scanning/ai-scan. It works like the existing pr_scan setting: an org can disable it for all repos, or enable it and let repos opt out. Turning it on doesn’t start a scan; users with write access still run scans on demand, and each scan uses GitHub Code Security AI credits. (changelog)

ARM32 retired on GitHub Actions hosted runners (12 Nov)

Hosted runners will end support for ARM32 job execution on Windows and Linux by 8 December 2026, including native AArch32 user-space execution. This follows the end of Node 20 support in Actions, because Node 24 doesn’t support ARM32. If any workflow still builds or tests 32-bit ARM on hosted runners, plan a move to self-hosted runners or emulation now. (changelog)

Secret scanning detects Cloudflare service tokens (12 Nov)

Secret scanning now detects cloudflare_access_service_token. Partner secrets found in public repositories are reported to the issuer automatically through the partnership program, and user secrets raise alerts in public and private repositories. (changelog)

Public monitoring alerts get direct links (12 Nov)

Each public monitoring alert in an enterprise’s Public monitoring view now has its own URL, so you can bookmark a finding or share it with teammates who have access. The feature is for GitHub Enterprise Cloud customers with Secret Protection or Advanced Security. (changelog)

How to use this list

  • Building an agent product? Look at AG-UI for the protocol and assistant-ui for the React front end.
  • Want your assistant to reach internal tools? Wrap them as MCP tools with FastMCP.
  • Analytics team? Try WrenAI on a read-only copy of your warehouse first.
  • Self-hosting fans: cloud-mail and pl_mail are two very different takes on owning your email.
  • Still on ARM32 hosted runners? You have until 8 December.

Takeaway

Today’s Trending page is about the infrastructure that makes agents usable: protocols, UI kits, MCP servers and shared skills, plus solid self-hosted and ML-core projects. Star the ones that fit your stack, read the READMEs critically and try one on a side project first. See you in tomorrow’s roundup.