#!/usr/bin/env node
/**
* Docfork MCP Server
*
* Main entry point supporting both stdio and HTTP transports.
* Automatically detects OpenAI clients and routes to appropriate server.
*/
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { CallToolResult } from "@modelcontextprotocol/sdk/types.js";
import { searchDocs, readUrl } from "./api/index.js";
import { z } from "zod";
import {
getServerConfig,
resolveAuthConfig,
setGlobalAuthConfig,
getAuthConfig,
} from "./config.js";
import { startHttpServer, startStdioServer } from "./server.js";
import { getServer as getOpenAIServer } from "./openai.js";
/**
* Create and configure the standard MCP server
*/
export const getServer = () => {
const server = new McpServer({
name: "Docfork",
version: "1.2.0",
websiteUrl: "https://docfork.com",
icons: [
{
src: "https://docfork.com/icon.svg",
mimeType: "image/svg+xml",
},
],
});
// register docfork search docs tool
server.registerTool(
"docfork_search_docs",
{
title: "Search Documentation",
description:
"Search documentation across GitHub repositories or the web. For targeted searches INSIDE a specific library's documentation, use the docforkIdentifier parameter (author/repo format). Extract from GitHub URLs (e.g., github.com/facebook/react → 'facebook/react') and include in ALL subsequent searches about that library for focused, accurate results.",
inputSchema: {
query: z.string().describe("Search query. Include language/framework names."),
docforkIdentifier: z
.string()
.optional()
.describe(
"CRITICAL for targeted library searches: Library identifier in author/repo format (e.g., 'facebook/react', 'vercel/next.js'). Use this to search INSIDE a specific library's documentation for focused, accurate results. Extract from URLs in docfork_search_docs results and ALWAYS include in all subsequent searches about that library."
),
tokens: z.string().optional().describe("Token budget: 'dynamic' or number (100-10000)"),
},
},
async ({ query, tokens, docforkIdentifier }): Promise<CallToolResult> => {
const authConfig = getAuthConfig();
const response = await searchDocs(
query as string,
docforkIdentifier as string | undefined,
tokens as string | undefined,
authConfig
);
return {
content: response.sections.map((section) => ({
type: "text" as const,
text: `title: ${section.title}\nurl: ${section.url}`,
})),
};
}
);
// register docfork read url tool
server.registerTool(
"docfork_read_url",
{
title: "Read Documentation URL",
description:
"Fetches and returns the full content of a documentation page as markdown. This is ESSENTIAL for getting complete, detailed information after searching. Always use this to read URLs from docfork_search_docs results to get the actual documentation content.",
inputSchema: {
url: z
.string()
.describe("Full URL to read. Use exact URLs from docfork_search_docs results."),
},
},
async (args): Promise<CallToolResult> => {
const inputValue = args.url as string;
const authConfig = getAuthConfig();
const response = await readUrl(inputValue, authConfig);
return {
content: [
{
type: "text" as const,
text: response.text,
},
],
};
}
);
// Error handler
server.server.onerror = (error: any) => {
console.error("MCP Server error:", error);
};
return server;
};
/**
* Main function
*/
async function main() {
const config = getServerConfig();
if (config.transport === "stdio") {
// resolve auth config from CLI args and env vars only (no headers for stdio)
try {
const authConfig = resolveAuthConfig();
setGlobalAuthConfig(authConfig);
} catch (error: any) {
console.error(`Configuration error: ${error.message}`);
process.exit(1);
}
await startStdioServer(getServer);
} else {
// HTTP transport with client detection
// auth config will be resolved per-request from headers and stored in AsyncLocalStorage
await startHttpServer(config.port, getServer, getOpenAIServer);
}
}
// Handle graceful shutdown
const shutdown = (signal: string) => {
console.error(`Received ${signal}, shutting down gracefully...`);
process.exit(0);
};
process.on("SIGINT", () => shutdown("SIGINT"));
process.on("SIGTERM", () => shutdown("SIGTERM"));
main().catch((error) => {
console.error("Fatal error running server:", error);
process.exit(1);
});
