Get An Expert (MCP server)

Stuck in an AI coding session? Get An Expert connects you with a real human expert.

This MCP server plugs into Claude Code, Codex, Cursor, Windsurf, VS Code, and other MCP-capable coding tools. When a session goes in circles, your coding agent can offer to bring in Get An Expert — and only with your explicit yes, it sends one redacted, structured summary of where you're stuck. That opens a live chat terminal where a human expert joins you — a direct, human-to-human conversation, like texting; no AI reads it. While the chat is open your session (prompts, agent replies, agent-run commands with output, file edits) relays live to the expert so they can watch real attempts — with a RELAY ON indicator, /pause, and /end always under your control.

Real humans: every response comes from a human expert — all types of experts, matched to the problem you're stuck on. No AI-generated answers.

Install

Claude Code

claude mcp add get-an-expert -- npx -y get-an-expert-mcp

Or add to your project's .mcp.json:

{
  "mcpServers": {
    "get-an-expert": {
      "command": "npx",
      "args": ["-y", "get-an-expert-mcp"]
    }
  }
}

There's also a companion Claude Code plugin that adds deterministic stuck-detection (a Stop hook that counts repeated failed attempts) on top of this server.

Codex CLI (~/.codex/config.toml)

[mcp_servers.get-an-expert]
command = "npx"
args = ["-y", "get-an-expert-mcp"]
startup_timeout_sec = 30  # first npx run downloads the package

Cursor (~/.cursor/mcp.json) / Windsurf (~/.codeium/windsurf/mcp_config.json)

{
  "mcpServers": {
    "get-an-expert": {
      "command": "npx",
      "args": ["-y", "get-an-expert-mcp"]
    }
  }
}

VS Code (.vscode/mcp.json — note the servers key)

{
  "servers": {
    "get-an-expert": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "get-an-expert-mcp"]
    }
  }
}

Replit

Replit supports remote MCP servers only; a hosted endpoint is on the roadmap.

How it works

1. You get stuck. Ten messages deep, same error, going in circles.
2. Your agent offers help. The offer_expert_help tool shows you exactly what would be sent, what is never sent, and asks for your yes/no. It transmits nothing.
3. You decide. Only after you agree does request_expert_help send one structured summary (goal, attempts, errors, short session summary, tech stack). Secret redaction runs on your machine first. In hosts that support MCP elicitation (Claude Code, Cursor, VS Code) you also get a native confirmation dialog.
4. A chat terminal opens and the relay arms. A window running npx get-an-expert chat <id> opens (or the command is printed for you). A human expert joins you there; while the chat is open, your session relays live to them so they see real attempts, not retellings. Every submission comes with a private deletion link.
5. Back-and-forth until unblocked. You try the fix in your normal session, the expert watches it land and advises in the chat. Either side can end it anytime — the instant it ends, nothing relays anymore. Afterwards, ask your agent to "apply what the expert suggested" and check_expert_replies pulls the advice back in.

Tools

| Tool | Network | What it does | |---|---|---| | offer_expert_help | none | Returns the offer + consent notice for the user | | request_expert_help | sends to Get An Expert API | After explicit consent, submits the summary, opens the live expert chat, and arms the session relay | | check_expert_replies | reads from Get An Expert API | Fetches the expert's new chat messages for this machine's most recent chat | | get_privacy_info | none | Returns the plain-words data-handling summary |

Configuration

| Env var | Default | Purpose | |---|---|---| | GET_AN_EXPERT_API_URL | https://get-an-expert.vercel.app | API base URL override |

Privacy Policy

Full policy: https://get-an-expert.vercel.app/privacy · Terms: https://get-an-expert.vercel.app/terms

• Collected (only when you explicitly agree, per request): your stated goal, what was tried, error messages, a short session summary, tech stack, and a random install ID (a UUID stored at ~/.get-an-expert/install-id, not linked to any account). Once you proceed, your chat messages and — only while the chat is open — relayed session activity (prompts, agent replies, agent-run commands with output, file edits) are collected under that same one-time consent; ending the chat is a hard stop the server enforces.
• Never collected: your source files, environment variables, or secrets. Client-side redaction runs before transmission (the summary, every chat message, every relayed event), and a second pass runs server-side. Zero bytes are sent before you consent, and nothing relays outside an open chat — no telemetry, no background transmission of any kind.
• Usage & storage: summaries, chat messages, and relayed events are seen by Get An Expert's human experts to help you, and stored with Upstash (Redis); hosting is on Vercel. No selling of data, no advertising use, no model training on your data.
• Retention: requests, their chat, and their relayed events auto-delete 30 days after submission. Every request includes a private deletion link that removes everything immediately.

License

MIT