跪拜 Guibai
← All articles
MCP · AI Coding · Agent

Protocols, Not Models, Will Decide Which AI Coding Tools Win

By 莪_幻尘 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Your next AI tool's real cost is protocol fit: whether it plugs into your existing MCP servers, streams state to your frontend over AG-UI, and carries your team's skills across tool switches. Ignore these layers and you'll pay in rewrites and vendor lock-in.

Summary

AI coding tool competition has moved from raw model capability to protocol compatibility. The Linux Foundation's Agentic AI Foundation now hosts Anthropic's MCP, OpenAI's AGENTS.md, and Block's goose under one neutral roof, with Amazon, Google, Microsoft, OpenAI, Anthropic, Cloudflare, and Salesforce on board. That institutional backing means investments in these protocols carry less single-vendor lock-in risk.

The protocol stack splits into four layers: MCP for tool invocation, A2A for agent collaboration, AG-UI for streaming process state to frontends, and Agent Skills for portable team norms. AG-UI defines about 27 event types and a frozen 1.0 spec, letting a single agent feed web, mobile, and terminal clients. Chrome DevTools MCP wraps CDP as MCP tools, so agents can navigate a live browser, read Console, inspect network, and run Lighthouse, closing the loop between editing code and seeing actual rendering. stitch-skills packages design constraints as installable skill files that work across Codex, Claude Code, Cursor, and others.

Three pitfalls remain: long-session browser connections become unstable, event streams desync without snapshot-plus-delta handling, and skill portability still requires manual folder copying in OpenCode. The practical test for any AI tool is now whether it can connect to your existing MCP servers, stream state over AG-UI, and load your team's skills.

Takeaways
— The Linux Foundation's Agentic AI Foundation hosts Anthropic's MCP, OpenAI's AGENTS.md, and Block's goose under one neutral platform, with Amazon, Google, Microsoft, OpenAI, Anthropic, Cloudflare, and Salesforce as members.
— AG-UI defines about 27 standard event types and froze its 1.0 spec; JSON Schema generates TypeScript, Python, and .NET SDKs, backward compatible with 0.x.
— AG-UI hit 17k-18k GitHub stars around November 2025, adding over 1,600 stars in a single week.
— Chrome DevTools MCP exposes navigation, clicks, snapshots, console, network, performance traces, and Lighthouse as MCP tools.
— Antigravity 2.0 bundles Chrome DevTools MCP; developers can invoke it via the /browser command or a sub-agent.
— Chrome DevTools MCP gained 717 stars on GitHub Trending on Nov 18, 2026, with a reported star count above 39K.
— stitch-skills is a Google Labs skill library with about 8.7K stars, compatible with Codex, Antigravity, Gemini CLI, Claude Code, Cursor, and OpenCode.
— OpenCode requires manually copying skill folders and enforces kebab-case naming, showing the last-mile standardization seams.
— Evaluation criteria for AI coding tools include MCP server compatibility, AG-UI state streaming, Agent Skills loading, runtime feedback (browser/console/network/performance), observability, and replaceability.
Conclusions

The institutionalization of competing AI standards under the Linux Foundation lowers the risk of single-vendor lock-in, a direct consequence of the agreement to share a table.

AG-UI's value lies in being a transport contract, not a UI framework; treating it as the latter guarantees disappointment.

Chrome DevTools MCP's debugging loop is orthogonal to Playwright's test-driven approach; one is about live iterative feedback, the other about scripted validation.

Protocols standardize data flow, while Agent Skills standardize judgment; both are necessary for an agent to work like a team, not just connect to systems.

Concepts & terms
MCP (Model Context Protocol)
A standard that defines how AI agents invoke external tools and data sources, decoupling tool capabilities from any single product.
AG-UI
An event-driven transport layer between AI agents and user interfaces; defines a typed, ordered event stream so frontends can render agent progress and state.
Agent Skills
An open standard for packaging reusable 'how-to' knowledge (design rules, review checklists, coding norms) into installable, versionable folders that any compatible agent can load.
Chrome DevTools MCP
An MCP server that wraps Chrome DevTools Protocol, giving coding agents live browser control, console inspection, network monitoring, and performance tracing.
stitch-skills
Google Labs' open-source skill library that encodes design constraints as portable skill files for UI agents.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗