Datapoint MCP

Get real human opinions from inside any MCP client. Run surveys, A/B preference comparisons, ratings, and rankings on text, images, audio, and video — without leaving your editor.

Datapoint MCP is an MCP server that gives Claude, GPT, Gemini, and any other MCP-capable agent the ability to recruit real humans for evaluation tasks, then return aggregated results back into the conversation. Built on top of Datapoint AI.

Why

LLMs are great at generating options and bad at telling you which one a real person will prefer. Datapoint MCP closes that loop — your agent can hand off to a panel of real humans and pick up the results a few minutes later.

Use cases

• Design & UX — A/B test logos, landing pages, screens, ad creative, copy
• AI evaluation — human ratings of model outputs, side-by-side comparisons, hallucination checks
• Preference data — collect RLHF / DPO pairs at scale
• Dataset labeling — classification, ranking, captioning, content moderation
• Product research — quick concept tests, naming, pricing reads
• Human-in-the-loop checks — gate an agent before it ships something irreversible

Tools

How it compares

Links

• Homepage: trydatapoint.com
• MCP spec: modelcontextprotocol.io
• Issues / discussions: GitHub

Install

Requires uv on your PATH.

Claude Code

As a plugin (recommended):

/plugin marketplace add impel-intelligence/datapoint-mcp
/plugin install datapoint@datapoint

To pick up new versions: /plugin marketplace update datapoint then /plugin update datapoint@datapoint.

As a raw MCP server (in ~/.claude/settings.json):

{
  "mcpServers": {
    "datapoint": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/impel-intelligence/datapoint-mcp.git", "datapoint-mcp"]
    }
  }
}

Claude Desktop

Add to claude_desktop_config.json (~/Library/Application Support/Claude/ on macOS, %APPDATA%\Claude\ on Windows):

{
  "mcpServers": {
    "datapoint": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/impel-intelligence/datapoint-mcp.git", "datapoint-mcp"]
    }
  }
}

Restart Claude Desktop, then ask it to run setup.

Cursor

Add to ~/.cursor/mcp.json (or via Settings → MCP):

{
  "mcpServers": {
    "datapoint": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/impel-intelligence/datapoint-mcp.git", "datapoint-mcp"]
    }
  }
}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "datapoint": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/impel-intelligence/datapoint-mcp.git", "datapoint-mcp"]
    }
  }
}

VS Code (GitHub Copilot Chat / agent mode)

Add to your workspace .vscode/mcp.json:

{
  "servers": {
    "datapoint": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/impel-intelligence/datapoint-mcp.git", "datapoint-mcp"]
    }
  }
}

Any other MCP client

Run the binary over stdio:

uvx --from git+https://github.com/impel-intelligence/datapoint-mcp.git datapoint-mcp

Usage

Once installed, just ask:

"Survey 20 people: which logo do they prefer, A or B?"

"Get human ratings on these three model outputs — which sounds most natural?"

"Run a quick A/B test on these two landing-page headlines."

The agent calls plan_survey to design it, shows you the plan and cost, then calls create_survey to launch. Use check_survey to monitor progress and read aggregated results.

Run setup first to authenticate if you haven't already.

Chain surveys (multi-step flow)

Some surveys have dependent questions — the second only makes sense given a specific answer to the first. Describe it that way and Claude will plan a chain:

"Ask 20 listeners if they could understand the speaker in this clip. If yes, rate the audio quality 1–5. If not, skip the rating."

A chain ties 2–5 steps together into a single unit of annotator work: every step is served to the same annotator, in order, and a per-step skip_if rule can end the response early. Claude will show you the full chain structure (steps, any skip conditions, cost) and wait for your confirmation before calling create_survey.

The cost shown in plan_survey is the upper bound (every response answers every step); when skip_if rules fire, responses cost proportionally less.

Configuration

License

MIT — see LICENSE.