MCP server

#Why MCP?

When you control the agent runtime, MCP is the lowest-friction way to give it new tools — no SDK install, no JSON-schema definitions, no per-tool wiring.

Drop one block of config into Claude Desktop or Cursor (or any conformant client) and Verdacert's seven tools show up in the tool picker. The agent calls them as if they were native. You get the same idempotency, rate limiting, structured errors, and signed certificates as the REST API — same backend.

#Available tools

Identical semantics to the REST endpoints; consult /docs/api for full request/response shapes.

#Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (or the Windows equivalent at %APPDATA%\\Claude\\claude_desktop_config.json) and restart Claude Desktop.

{
  "mcpServers": {
    "verdacert": {
      "url": "https://verdacert.com/api/mcp",
      "headers": {
        "Authorization": "Bearer vc_sandbox_…"
      }
    }
  }
}

Confirm the integration by asking Claude something like: “Can you call Verdacert and tell me which languages you can certify translations for?” Claude should call get_capabilities automatically.

#Cursor

Settings → MCP → Add new MCP server. Paste the same JSON shown above. Save.

The tools show up in Cursor's in-chat tool picker immediately. Test the connection by typing /verdacert in chat to surface the tools.

#Vercel AI SDK (MCP client)

The AI SDK has a built-in MCP client if you'd rather consume MCP than install our typed tools package.

import { generateText, experimental_createMCPClient } from "ai";

const mcp = await experimental_createMCPClient({
  transport: {
    type: "sse",
    url:  "https://verdacert.com/api/mcp",
    headers: {
      Authorization: `Bearer ${process.env.VERDACERT_API_KEY}`,
    },
  },
});

const tools = await mcp.tools();

const { text } = await generateText({
  model: "anthropic/claude-sonnet-4",
  tools,
  prompt: "Quote a 3-page Farsi birth certificate for USCIS.",
});

await mcp.close();

#Raw JSON-RPC

If you're integrating from a runtime without an MCP library, the wire format is plain JSON-RPC 2.0 over POST.

curl https://verdacert.com/api/mcp \
  -H 'Authorization: Bearer vc_sandbox_…' \
  -H 'Accept: application/json, text/event-stream' \
  -H 'Content-Type: application/json' \
  -d '{
    "jsonrpc": "2.0",
    "id":      1,
    "method":  "tools/list",
    "params":  {}
  }'

curl https://verdacert.com/api/mcp \
  -H 'Authorization: Bearer vc_sandbox_…' \
  -H 'Accept: application/json, text/event-stream' \
  -H 'Content-Type: application/json' \
  -d '{
    "jsonrpc": "2.0",
    "id":      2,
    "method":  "tools/call",
    "params": {
      "name":      "quote",
      "arguments": {
        "sourceLanguage": "fa",
        "useCase":        "uscis",
        "pageCount":      2,
        "speedTier":      "standard"
      }
    }
  }'

#Behavior & limits

What an MCP-using agent sees vs. a REST client.

• Stateless. No session id; each JSON-RPC request stands alone. Combined with enableJsonResponse: true, every response is a single JSON body — no SSE stream to manage.
• Authenticated once per request. The same Bearer header authenticates each call; revoke a key and MCP access stops within the request.
• Rate limited. Same 300 req/min per-key limit as REST. Tool calls inside a single agent loop count individually.
• Structured errors bubble up through the MCP tool-result channel — agents see the same code / message / recoveryHint envelope and can self-correct.
• Max duration 60s per HTTP call. Real translation work happens out-of-band; each tool call returns quickly.

Where to next