Why does your AI still respond in markdown?

Outkit processes the completed response — it does not intercept your stream mid-flight. Once your LLM finishes generating, Outkit classifies the output and immediately begins streaming back text and rich components. Your existing markdown can remain visible to users while Outkit processes, so there's no blank screen or loading state.

Nothing bad. When confidence is low, Outkit falls back to your original text — your user sees exactly what they would have seen without Outkit. That request is not charged. We optimize against wrong upgrades because a missed enhancement is invisible, but a bad one breaks trust.

We never see your LLM API keys. We never see your prompts, conversation history, or message context — because we're post-processing, not proxying. We only receive the final output text you send us. Processing data is retained for debugging purposes within your account and design profiles. We also offer a zero-retention mode for teams with strict data handling requirements.

React and Next.js today, with Vue.js coming soon. The core classification API is framework-agnostic — it returns JSON, so you can call it from any language or backend. The client packages are thin wrappers around a self-contained Preact runtime in Shadow DOM, so they work without conflicting with your existing setup.

We skip it. If the response is pure prose with nothing worth componentizing, the classifier returns a signal and your user sees standard text. That request is not charged. Outkit only activates when there's something meaningful to upgrade.

Those approaches require your agent to know about UI at inference time — component schemas occupying context, tool definitions for layout and rendering, skills for generating frontend code, extra reasoning tokens spent on presentation instead of the user's actual question. All of that lives inside your agent loop, affecting every part of your architecture: context windows, tool routing, token costs, and agent performance. Outkit sits entirely outside. Your agent finishes its job, we handle how it looks. Zero impact on any part of your AI product architecture.

The initial classification adds a brief processing step after your response is complete, after which Outkit immediately starts streaming output with text and rich components. Net perceived latency is zero if your markdown is already visible to your users while Outkit processes — they're reading content the entire time.

Yes. Outkit ships as both an API and an MCP tool. Your agent can call outkit_enhance as an MCP tool to classify and render responses, or you can call the API directly from your backend. Either path, same result.