Docs · MCP servers

The Model Context Protocol is the open standard for connecting agents to tools and data. Your agent runs an MCP client; servers expose tools over stdio or HTTP; the client converts them to OpenAI function format and the loop above just works. Unrefused serves the brain.

The ecosystem has consolidated fast: the official MCP Registry passed 45,000 listings in late 2026, with directories tracking 150,000+ servers. Quality varies wildly — these are the ones with real adoption and real benchmark track records.

The curated directory

Server	Org	What it gives your agent	Stars
`github/github-mcp-server`	GitHub	Repos, issues, PRs, Actions — natural-language repository management. Read-only and toolset-filtered modes.	33.7k
`microsoft/playwright-mcp`	Microsoft	Browser automation over the accessibility tree — the most-downloaded MCP server (28.6M npm downloads/mo).	38.4k
`ChromeDevTools/chrome-devtools-mcp`	Chrome	Performance traces, network inspection, live DOM debugging.	53.8k
`upstash/context7`	Upstash	Up-to-date library docs injected at call time — kills stale-API hallucinations.	63.3k
`browser-use/browser-use`	Browser Use	Full browser control: forms, extraction, bookings.	118k
`microsoft/markitdown`	Microsoft	Office/PDF/image files → markdown an agent can read.	190k
`netdata/netdata`	Netdata	Live infra metrics, logs, alerts with anomaly detection.	81k
`modelcontextprotocol/server-*`	MCP reference	Filesystem, fetch, memory, sequential-thinking — the maintained reference set.	reference

Stars measure adoption, not safety. For anything credentialled, prefer servers with scoped auth and start read-only.

Connecting one

typescript

import { Client } from "@modelcontextprotocol/sdk/client/index.js";
import { StdioClientTransport } from "@modelcontextprotocol/sdk/client/stdio.js";

const mcp = new Client({ name: "unrefused-agent", version: "1.0" });
await mcp.connect(new StdioClientTransport({
  command: "npx",
  args: ["-y", "@modelcontextprotocol/server-filesystem", "./workspace"],
}));

const { tools: mcpTools } = await mcp.listTools();
const tools = mcpTools.map((t) => ({
  type: "function",
  function: { name: t.name, description: t.description, parameters: t.inputSchema },
}));
// then: the standard tool-calling loop from /docs/agents

HTTP (Streamable) servers connect the same way with StreamableHTTPClientTransport and auth headers. The official registry is the safe place to discover servers — every listing names its transport and package.

What the benchmarks say

MCP tool use is hard, and the 2025–26 benchmark wave quantified it:

Benchmark	Scope	Hardest result
MCPMark	127 CRUD-heavy tasks across GitHub, Notion, Filesystem, PostgreSQL, Playwright — programmatic verification	Best model 52.56% pass@1; average task needs 16.2 turns and 17.4 tool calls
MCP-Universe	6 domains, 11 real servers, execution-based evaluators	Best model 43.72% (GPT-5); Grok-4 33.3%, Claude-4.0-Sonnet 29.4%
MCP-AgentBench	33 servers, 188 tools, 600 queries across complexity tiers	Outcome-oriented MCP-Eval methodology

Read that honestly: every frontier model fails most MCP tasks. The differentiators the benchmarks isolate are exactly what this platform sells — long-context tolerance for sprawling traces, unfamiliar-tool reasoning, and not losing turns to refusals mid-loop. Pick your servers from the directory, wire the loop from Agents, and eval before you scale.