An MCP server is a small service that your AI assistant can call during a conversation to fetch real data. Instead of guessing from training data, the AI queries your MCP server, reads the results, and answers based on what it finds. That one change makes answers accurate, specific, and up to date.
How an MCP Server Actually Works
The Model Context Protocol is a standard interface. It defines how an AI assistant discovers tools, calls them, and uses their output. The AI decides when to call a tool based on the conversation. You do not wire up anything manually.
A real interaction looks like this:
- You ask: "How do I configure webhooks in your API?"
- The AI calls your MCP server with a search query
- Your server returns the most relevant documentation pages
- The AI reads those pages and writes an answer citing your actual docs
The AI never loads your full documentation at once. It gets back a ranked list of relevant excerpts, which keeps the response focused and the context window small.
Here is the key point: the AI is not summarizing from memory. It is reading fresh content every time.
What Yavy's MCP Server Exposes
Yavy exposes a single tool called search_docs. It accepts a natural language query and returns the top matching pages from your knowledge base.
Each result includes:
- Page title and URL
- A relevant excerpt
- A relevance score
- The source project the page came from
You can connect multiple projects to one MCP server. If you have separate knowledge bases for your API docs and your help center, the AI can search both in the same conversation without any extra configuration.
One server. Multiple knowledge bases. The AI figures out which content is relevant.
Setting Up the MCP Server for Claude Code and Claude Desktop
Yavy generates a unique MCP server URL for each project. You add that URL to your AI assistant's configuration file.
For Claude Code, add it to .mcp.json in your project root:
{
"mcpServers": {
"yavy-docs": {
"url": "https://yavy.dev/mcp/your-project-token"
}
}
}For Claude Desktop, the same block goes in claude_desktop_config.json. The format is identical.
Restart your AI assistant after adding the config. It discovers available tools automatically on startup. No extra prompting required.
Keeping Documentation Results Fresh
Yavy re-indexes your documentation on a schedule you control. Daily re-indexing covers most teams. If you push documentation updates frequently, you can trigger a manual re-index from the dashboard or via the Yavy API.
The MCP server always queries the latest index. There is no cache to clear and no configuration to update when your docs change. Push a docs update, trigger a re-index, and the AI picks it up on the next query.
MCP Server vs. Skills Packages
Both approaches give your AI assistant access to your documentation. The difference is delivery.
MCP servers make a live network call for every query. That means results are always current, but it requires an internet connection and adds roughly one to three seconds to each response when a tool is called.
Skills packages bundle your documentation locally. The AI reads from a file rather than calling a server. This works offline and adds zero latency, but you need to update the package manually when your docs change.
The right choice depends on your workflow. Teams that update documentation regularly and want zero maintenance overhead should use MCP. Teams distributing a product offline, or those who need sub-second responses without network dependency, should use Skills.
Neither is universally better. The quality of answers is the same either way.
One practical note: the AI calls your MCP server when it judges that your documentation has relevant information. It does not call it for every message. For generic questions, it answers from training data instead.
You can override this with an explicit instruction: "Search the docs for how to set up OAuth." That forces a lookup even if the AI thinks it already knows the answer.
Two other common issues:
The AI says it cannot find information in your docs. Check that the MCP server URL is correct and that the project has finished indexing. Test the search directly from the Yavy dashboard to confirm results are returning.
Slow responses. MCP adds one network round-trip when a tool is called. If that latency is a problem, Skills packages eliminate it entirely.