TriFrost now features native support for the Model Context Protocol (MCP), and we've deployed it directly to trifrost.dev.
AI assistants like Claude and Cursor have transformed how code is written, but fetching up-to-date documentation often requires manual searches or relying on stale pre-training data. MCP solves this by standardizing the flow between AI tools and data sources.
A Stateless Approach
Most MCP servers run locally over stdio. Web-based MCP servers require Server-Sent Events (SSE).
Because Serverless platforms (like Cloudflare Workers) do not guarantee sticky sessions, an HTTP POST to /mcp/messages is not guaranteed to hit the exact V8 isolate running the long-polling /mcp/sse connection.
TriFrost circumvents this using native framework primitives:
1. ctx.file({stream}): Keeps the SSE socket open indefinitely.
2. ctx.cache: Leveraging DurableObjectCache, the stateless POST handler blindly writes JSON-RPC payloads into the distributed cache. The SSE isolate routinely polls this cache, consumes the messages, tracks execution via @modelcontextprotocol/sdk, and pushes responses back down the wire.
No custom Durable Objects. No complicated WebSocket connection logic.
What This Means
You can now point your local AI assistant directly at https://trifrost.dev/mcp/sse using an inspector proxy.
Your assistant can natively call:
get_doc_by_sluglist_examplesget_update_by_slug
When you ask your AI assistant "How do I build a TriFrost application?", it will no longer guess. It will ask TriFrost.
As always, stay frosty. ❄️