Ycode, a visual web-building platform, [shipped an AI agent](https://docs.ycode.com/docs/editor/ai) that does something unusual for the no-code space: it runs on the user’s own API key from Anthropic, OpenAI, Google Gemini, or xAI. Ycode does not charge for AI usage. It does not mark up tokens. It routes the chat through whichever provider the user picks and lets the billing fall where it may.
That is a meaningful break from the rest of the market. Every major no-code and low-code platform — Webflow, Wix, Squarespace, Bubble — has added AI features in the last two years. Almost all of them bundle the cost into the subscription or charge a separate per-credit fee. The user never sees the underlying model or its price. Ycode is betting that a segment of builders prefers to pay their provider directly and keep control of which model they use, rather than accept whatever the platform negotiates on their behalf.
The agent itself is a chat panel inside the Ycode editor. It can create and edit pages, manage CMS collections, modify layers and styles, add animations, handle localization, and work with components. It uses about 130 tools from Ycode’s MCP server, minus the publish action — the agent never ships anything automatically. Every change stays in draft until the user hits Publish. Each turn has a one-click Undo that reverts every change from that message.
The design system awareness is worth noting. Before every turn, the agent receives the project’s color variables, fonts, components, and shared Layer Styles as context. New sections use existing tokens rather than inventing new ones. The agent follows Ycode’s structural conventions — section, container, content — and runs a design-lint pass before finishing. This is a concrete attempt to solve the “looks like AI” problem that plagues generated websites. Whether it succeeds depends on how tightly the lint rules are written, but the architecture is sound: the agent is constrained by the same design system a human would use.
The model selection is unusually broad. Ycode ships a curated list per provider: Claude Opus 5 (the default), Claude Fable 5, and Claude Sonnet 5 on Anthropic; GPT-5.5 and GPT-5 Mini on OpenAI; the latest Gemini 2.x models on Google; and Grok 4.5 and Grok 4.3 on xAI. Users can pick a default model and enable or disable specific ones. The model picker is available in the chat composer, and switching models mid-conversation only affects subsequent turns. Session history records which model produced each reply, so revisiting a chat shows exactly which model wrote what.
The bring-your-own-key model has implications beyond pricing. It means Ycode does not need to negotiate enterprise agreements with model providers. It does not need to provision inference capacity. It does not need to handle the data-processing agreements that come with routing user content through a third-party API. The user’s data goes directly from the browser to the provider’s endpoint. Ycode’s infrastructure never sees the prompt or the response. That simplifies compliance for teams that work with sensitive content, and it eliminates a class of privacy concerns that have slowed AI adoption in regulated industries.
The tradeoff is complexity. Setting up the agent requires navigating to Settings, pasting an API key, testing it, choosing a key scope (shared with the project or personal), and selecting which models to enable. The docs are clear, but the friction is real compared to a platform that just turns on AI with a toggle. Ycode is betting that its target user — likely a solo builder or a small team that already uses Claude or ChatGPT for other work — will find the setup worth it for the control.
The agent tab has a master toggle in settings that can disable it entirely. When off, the tab is hidden and the chat API refuses requests. That is a small detail but a telling one. It signals that Ycode treats the AI agent as a feature that some projects will want to lock down, not as a core part of the editing experience that must always be on.
The broader context: AI coding agents are proliferating fast. Agentic.ai’s ranking of the top 20 coding agents in 2026 shows Claude Code and Cursor tied at the top with an agenticness score of 21 out of 36, followed by GitHub Copilot, Cline, and Devin Desktop at 19. The market is fragmenting by deployment model — terminal agents, IDE extensions, cloud-hosted agents, and now platform-embedded agents like Ycode’s. The no-code segment has been slower to adopt agentic workflows than the developer-tools segment, partly because the abstraction layers are thicker and the tool surface is wider. Ycode’s approach, using an MCP server to expose the full tool registry to the agent, is a clean architectural solution. The agent does not need special integration for each action. It calls tools from the same catalog that external clients like Cursor and Claude Desktop use over MCP.
The question that remains is whether the BYOK model scales beyond early adopters. Most no-code users do not have an Anthropic API key. They do not know what a token costs. They want to build a website, not manage API billing. Ycode’s docs address this by letting the shared key option cover an entire project, so one person on the team can set it up and everyone else just uses the agent. But the first mile is still a barrier.
What is notable is that Ycode chose this path at all. The standard platform play is to bundle AI, charge a margin, and abstract the provider away. Ycode is doing the opposite. It is treating the AI model as infrastructure that the user already pays for elsewhere, and positioning itself as the interface layer that makes that model useful for web design. That is a bet on interoperability over lock-in, and on user choice over platform convenience. The market will decide whether that bet pays off, but for now it is the most interesting pricing model in the no-code AI space.