← Back to the AI-Radar overview

MCP and agent tooling: what's happening now

Published on July 23, 2026 · Analysis for Indie Developers

The world of AI agents and the Model Context Protocol (MCP) is evolving at a breakneck pace. For indie developers running their own AI infrastructure—whether through a local LLM gateway like LiteLLM, a custom agent setup on a home server, or workflows in n8n—the latest wave of updates brings both challenges and massive opportunities. The most significant shift is the upcoming radical overhaul of the MCP protocol itself, which is transitioning to a stateless architecture with mandatory authentication. In addition, we are seeing powerful new open-source coding models running locally on consumer hardware, terminal-first tools competing with proprietary alternatives, and experimental decentralized frameworks.

In this article, we highlight the key signals from the community and explain what this means for your local AI stack.

The Great MCP Shift: Stateless and Mandatory Authentication

Signal: High Action: Update
MCP 2026-07-28 Protocol Release

The specification for the largest MCP update since its launch will be finalized on July 28, 2026. The most significant change is that the protocol will become completely stateless: the initialize handshake and session IDs will disappear. In addition, OAuth/OIDC authentication will be made mandatory. Other additions include MCP Apps (HTML in sandboxed iframes), a Tasks extension for long-running background tasks, and built-in support for OpenTelemetry. The update introduces about six breaking changes, with a 12-month migration window. Beta SDKs for Python, TypeScript, Go, and C# are now available.

Why this matters: If you are running custom-built or locally hosted MCP servers in your homelab, you need to prepare them for this transition. The introduction of mandatory authentication means you will have to tighten the security of your local endpoints, but it also makes it safer to expose your local servers to external agents. For smart routing of your local and external models, check out our guide on model routing on llmnet.

Source post by @kkaminsk

FastMCP 4.0 Alpha: Preparing for the New Spec

Signal: High Action: Test
FastMCP 4.0 Alpha release

Following the protocol update, Jeremiah Lowin (PrefectIO) has released the first alpha version of FastMCP 4.0. This version is built entirely on the new MCP SDK v2 specification for the 2026-07-28 release. Although there are some breaking changes, the goal is for most existing FastMCP servers to be able to upgrade without major modifications.

Why this matters: FastMCP is one of the most widely used Python libraries for quickly building MCP servers. If you have written custom integrations to control, for example, your local database or home server functions via an LLM, it is wise to start testing with this alpha now to prevent your workflows from breaking once the final spec goes live.

Source post by @jlowin

KAT-Coder-V2.5-Dev: Powerful Local Coding on Consumer Hardware

Signal: High Action: Implement
KAT-Coder-V2.5-Dev open-source 35B MoE model

KwaiKAT has released KAT-Coder-V2.5-Dev, an open-source (Apache 2.0) Mixture of Experts (MoE) model with 35 billion parameters (of which 3 billion are active per token). The model is post-trained on Qwen3.6 and achieves a score of 69.4% on the SWE-bench Verified benchmark. With a context window of 262k tokens, native tool use, and a special 'thinking mode', this model is specifically designed for agentic coding. It runs with llama.cpp at a speed of about 89 tokens per second on a dual RTX 5060 + 4060 Ti setup (32GB VRAM).

Why this matters: This model offers a serious, privacy-friendly alternative to expensive, cloud-based APIs like Claude or OpenAI for programming tasks. Because it runs efficiently on relatively affordable consumer GPUs, it is an ideal candidate for hosting in your own homelab via tools like Ollama or vLLM.

Source post by @TraffAlex

Zero v0.5.0: A Terminal-First Alternative to Claude Code

Signal: Medium Action: Evaluate
Zero v0.5.0 Go terminal-first agentic coding tool

Zero is an open-source, Go-written CLI tool for agentic coding directly from your terminal. With the release of version 0.5.0, the project introduces features such as concurrent read-only batching, prompt caching, a unified sandbox environment, an /undo command, and /btw for side conversations with the agent without cluttering the main context. The project now has over 1,100 stars on GitHub.

Why this matters: If you are looking for an open-source and locally runnable alternative to tools like Claude Code or GitHub Copilot CLI, Zero offers a powerful workflow. It integrates seamlessly into a terminal-based development environment and allows you to let agents work directly on your codebase without vendor lock-in.

Source post by @EngMoElgaraihy

Buzz: Decentralized Agent Orchestration via Nostr

Signal: Medium Action: Explore
Buzz: decentralized AI agent workspace on Nostr

Buzz, developed by Jack Dorsey's Block, is an open-source (Apache 2.0) workspace where humans and AI agents collaborate as peers. It is built entirely on the Nostr protocol. Every action within the system is a digitally signed Nostr event. The platform includes an agent harness (buzz-acp) that acts as a bridge to LLMs like Claude, Goose, or local instances via vLLM and Ollama. Additionally, it offers encrypted agent memory (known as 'engrams') and 'owner attestations' for fine-grained authorization. The codebase now comprises around 300,000 lines of Rust.

Why this matters: Although this project is still experimental, it paints a fascinating future for multi-agent systems. By using Nostr as a communication layer, Buzz solves problems around cryptographic provenance and agent identity. For developers interested in decentralized architectures and sovereign AI systems, this is a project to follow closely.

Source post by @clawement

Codebase Mapping with Atlas Scout

Alongside these open-source developments, there is Atlas Scout (by Zaguan AI). This is a local, read-only MCP server specifically designed to provide coding agents with a structural map of a codebase. This prevents agents from searching "blindly" with commands like grep, saving valuable context and tokens. Although the product is closed-source, it offers a free entry-level tier and a trial period. It demonstrates how specialized the tooling around MCP is becoming. Even if you prefer fully open-source solutions, Atlas Scout shows that a good "codebase map" is crucial for local agents to function efficiently.

What can you do with this?

As an indie developer with your own AI stack, you can act on these signals immediately: