15.1 Where MCP Is Heading

MCP (Model Context Protocol) began as a way to let an AI agent call an external tool. Sixteen months later, it has become infrastructure --- and the 2026-07-28 RC is the largest revision since the protocol launched in November 2024.

As of June 2026, the MCP specification is stable at 2025-11-25. The 2026-07-28 RC, scheduled to finalize on July 28, 2026, introduces four changes that matter for Open Design specifically.

Stateless HTTP. The current protocol requires session affinity --- a client and server must hold a persistent connection. The RC makes the core protocol stateless: clients cache the tools list, servers run behind ordinary HTTP load balancers. This means MCP can operate at cloud scale without custom session routing. For local-first use, it changes nothing. For the ecosystem, it removes the biggest adoption barrier for enterprise deployments. That matters to Open Design because it means the MCP surface that Open Design exposes will be reachable from more agents running in more environments.

MCP Apps. The RC adds a mechanism for servers to ship interactive HTML into sandboxed iframes with bidirectional JSON-RPC. In plain terms: an MCP server can now render a UI panel inside the agent's host application. This is structurally similar to what Open Design already does with its preview iframe --- the difference is that MCP Apps puts that capability in the protocol itself, so any MCP-compatible host can surface it without custom integration code.

Tasks. Long-running work moves to an independent extension. Generation runs in Open Design are already long-running; this extension standardizes how agents report progress and intermediate results across the ecosystem.

OAuth 2.0 / OIDC with Resource Indicators (RFC 8707). Standardized auth prevents token misuse across multi-server setups. This matters most in team and enterprise deployments of Open Design.

By March 2026, MCP hit 97 million monthly SDK downloads, up from approximately 2 million a year earlier. It is used across Claude Code, ChatGPT, Gemini, GitHub Copilot, and Cursor. The MCP Registry transferred to the Linux Foundation's Agentic AI Foundation on March 20, 2026. That governance structure --- an open foundation, vendor-neutral, with a chartered roadmap --- is the best available signal that MCP is infrastructure and not a vendor play.

15.2 The Agent and Model Trajectory

The agent market is in its most volatile phase. Between April and June 2026, a major model pricing cut and the sunset of an entire agent CLI happened within weeks of each other. That pace is the norm, not an anomaly.

The Consolidation Timeline

The clearest example of agent volatility is Gemini CLI. Google announced its sunset on the Google Developers Blog, with the service stopping for consumer tiers on June 18, 2026. According to multiple reports including The Register and the Google Developers Blog announcement, its announced replacement is Antigravity CLI, described in those reports as closed-source and Go-based, in contrast to Gemini CLI's open-source model. Multiple news sources and community discussions noted that this transition was announced without a prior open-source community consultation process. That is the documented cost of anchoring your workflow to a single vendor's CLI.

At the same time, OpenCode has emerged as the leading open-source alternative, reaching approximately 150,000 GitHub stars and a reported 6.5 million monthly active developers as of Q2 2026, per OpenCode's own published figures (opencode.ai). It is MIT-licensed, model-agnostic, and supports 75+ LLM providers. Those numbers will continue to grow; the point is not the count but what they signal: there is now a credible, community-owned option at the terminal-native tier.

The agent market has consolidated into five observable categories as of June 2026:

Model Cost Trajectory

Between February 2026 and June 2026, Anthropic reduced Claude Opus 4.7 pricing from the previous Opus 4.1 rate of $15/$75 to $5/$25 per million tokens --- a reduction of approximately 67% in four months. As of June 2026, GPT-5.5 is reported at $5/$30 per million tokens, with a $0.50 cached-input option for repeated context. Small and medium models (Haiku-class, GPT mini, Gemini Flash) have seen even steeper reductions since 2023 --- one pricing tracker (morphllm.com) puts the cumulative drop at 99.7% from the 2023 baseline, though that figure is from a single source and should be treated as directionally correct rather than precise.

The practical implication: the cost barrier to running local-first agent workflows has collapsed. Design generation requires multiple model calls per artifact --- the initial generation pass, iteration rounds, export validation. At $5 per million input tokens, a complete generate-preview-iterate-export loop costs less than a cup of coffee. The durable layer --- DESIGN.md, the local daemon, the skill catalog --- becomes more valuable as the cost of running inference against it approaches zero.

There is also a second-order effect worth naming: as inference costs fall, the differentiator is no longer which model you use, but how well you have structured your design intent. A DESIGN.md that precisely encodes your brand's type ramp, spacing scale, and component rules will consistently produce better output than a vague prompt regardless of the underlying model. Cheap inference raises the relative value of good brand contracts.

Consolidation is not over. Cognition Labs acquired Windsurf (formerly Codeium's IDE) and relaunched it as Devin Desktop in 2026. Near-weekly Codex CLI releases mark OpenAI's terminal-native push. The five-category taxonomy above describes where things stand in June 2026, but it will likely need revision by end of year --- probably because one or more categories collapses through acquisition.

15.3 Design-as-Code Deepens

The DESIGN.md specification is the most consequential recent development for local-first autonomous design. Not because it is new --- Open Design has used design-as-code contracts since its first release --- but because Google open-sourced the specification on April 21, 2026, which means the format is now standardized, portable, and no longer tied to any one tool.

The DESIGN.md spec, published at github.com/google-labs-code/design.md as of April 21, 2026, defines a machine-readable brand contract structured around W3C Design Token Community Group (DTCG) format. Tokens export to tokens.json, tailwind.config.js, and design-system.css natively. The spec is documented as compatible with Claude Code, Cursor, OpenCode, and Google's own Stitch tool.

Within 10 days of the associated "awesome list" being published in March 2026, it reached 40,500+ GitHub stars --- among the fastest initial growth of any awesome list on record, according to one tracking source. That speed of adoption is a signal that the design-as-code model was already latent demand waiting for a portable specification to anchor it.

This matters for Open Design specifically because the DESIGN.md files in Open Design's bundled brand systems are now written against a portable, spec-compliant format. If Open Design ceased to exist tomorrow, every DESIGN.md in your project directory would still work with any agent CLI that implements the spec. The durable layer, volatile layer model is not a metaphor --- it is a practical portability guarantee.

15.4 Open-Source Community Dynamics

Open Design at v0.9.0, released June 2, 2026, carries 63,000+ GitHub stars and 7,000+ forks. Ten releases in two months (April through June 2026) point to a monthly release cadence and an active maintainer group. By the time you read this, v1.0.0 is likely out.

These counts age weekly and should not be treated as precise --- the point is the pattern. Monthly releases, consistent star growth, and approximately 310 contributors (inferred from search results; ) describe a project in active maintenance, not a project coasting on early momentum.

The Open Design Fellow program offers funded community engagement, including reported incentives for merged pull requests and LLM credits for contributors. The strategic intent is clear --- lower the barrier to building skills and plugins, so the catalog keeps growing without the core team having to own everything.

More broadly, the design tool market is unsettled. A 2026 Figma design trends report noted that nearly half of designers are still searching for their go-to stack, and the average toolstack has doubled to seven tools. Local-first autonomous design is one answer to that fragmentation, not the only one. The community dynamics of Open Design --- an Apache-2.0 license, a funded Fellow program, an active Discord, and a plugin ecosystem that grows with each release --- are a hedge against the project itself stalling. If the core team slows down, the plugin and skill contributors carry it forward. That is the structural advantage of an open-source project with a large enough contributor base, and it is one of the reasons I read the 7,000+ fork count as a signal of genuine community investment rather than passive curiosity.

15.5 My Bet on What Endures

Forward-looking chapters are cheaply written and expensively wrong. I will be specific about what I think is durable, what I think is volatile, and where I could be wrong on each.

The durable layer, volatile layer model is the mental model this chapter has used throughout, and it is also my actual operating framework. Here is how I apply it to a concrete investment decision.

The phrase to hold onto is durable layer, volatile layer. Build everything repeatable --- your brand contract, your skill library, your export workflows --- on the durable side. Treat agents, models, and pricing as infrastructure you rent, not infrastructure you own. That posture keeps you competitive regardless of which tool wins the next eighteen months.