Read RPC docs from Node.js with mcp-use—without a model or API key

You can retrieve RPC setup documentation through MCP without starting a model-driven agent. This Node.js example lists the available documentation tools, searches for quickstart, and prints the Markdown returned by read_doc. It reads instructions; it does not execute the RPC or account steps inside them.

The endpoint is BlockVectra's documentation-only https://docs.blockvectra.com/mcp/docs, served over Streamable HTTP. It exposes only search_docs and read_doc. No BlockVectra account, API key, wallet, model API, or local MCP server is needed.

You need Git, npm, and Node.js 22.22.2 or newer. Start from this exact revision so the code and committed dependency lockfile match:

git clone https://github.com/blockvectra/agent-quickstart.git
cd agent-quickstart
git checkout --detach c4c0d5c81c5518add7f0f204a77f42934d183588
cd examples/mcp-use-docs
npm ci --ignore-scripts
npm start

The example pins @mcp-use/client to 2.4.5. Installation needs network access to the npm registry; it happens before the script starts. npm start runs node remote-docs.ts. The example disables SDK anonymous telemetry before importing the client module:

process.env.MCP_USE_ANONYMIZED_TELEMETRY = "false";
const { MCPClient } = await import("@mcp-use/client");

The first output line lists the two documentation tools. The script then searches in English with a one-result limit:

    const search = await connection.callTool(
      "search_docs",
      { query: "quickstart", lang: LANG, limit: 1 },
      { timeout: TIMEOUT }
    );

Here, LANG is "en" and TIMEOUT is 15000 milliseconds. The script parses the search result's text as JSON and checks that its first result contains a usable path. It passes that actual returned value to read_doc with the same language—not a document ID, a full URL, or a guessed filename.

Look for first result: followed by the selected path, then the page's Markdown. Empty or malformed search results stop the script rather than silently substituting a page. Tool errors and missing text also fail the run; the client closes in finally on success or failure. Calls have 15-second timeouts.

A November 11 internal check using Node.js 26.8.1 recorded four HTTP 200 responses for discovery, tool listing, search, and read; that is one tested configuration, not a performance benchmark or a model-agent walkthrough.

The complete source keeps the connection and error handling together. Documentation success does not establish business RPC or Data API success: those use the full connection and their own public-method or authentication policies.

For client setup, see the MCP server guide.