人口智慧 · MCP

把模擬焦點團體，
變成 AI 代理可呼叫的工具。

AI 代理可在任務進行中透過 CrowdOS MCP 伺服器發起調查，包括 Free 在內的每個方案均可使用。加入一個 URL，或執行一次 pip install。樣本組依人口普查資料抽取，內部意見相左，而不是單一模型的意見。

MODEL CONTEXT PROTOCOL·CLAUDE DESKTOP · CURSOR · CLINE·15 個可供代理呼叫的工具·PRO 與 MAX 已包含

什麼是 MCP

Claude 早就在說的協定

Model Context Protocol（模型脈絡協定）是 Anthropic 提出的開放標準，用來把 AI 代理連接到外部服務。Claude Desktop、Cursor、Cline、Zed，以及愈來愈多的代理框架（LangGraph、CrewAI、AutoGPT、Devin）都原生支援它。把一個 MCP 伺服器寫進宿主程式的設定檔，它的工具就會出現在模型的工具選擇器裡——不需要膠水程式碼，也不必為每個應用程式各接一套 SDK。

CrowdOS 提供兩種接入方式：由 API 直接託管的端點，以及給桌面宿主程式使用、可用 pip 安裝並中繼到該端點的 stdio 伺服器。兩者為你的帳戶提供相同的工具，並負責請求整形、回應精簡與錯誤去敏。

HOSTS

LangGraph and CrewAI run the agent. CrowdOS is the crowd.

CrowdOS is the crowd an agent should query. LangGraph and CrewAI are hosts: the place the agent runs. Mount the server and the same tools show up in the model's tool picker. AutoGPT and Devin speak the protocol too, the same way Claude Desktop, Cursor and Cline do.

A PROMPT IS NOT A PANEL

An agent role-playing a panel isn't enough

One model acting like N personas (effectively n=1).

An LLM persona is a costume for one prompt. Ours is a database row: census-grounded people with a fixed identity, an opinion history, and a standing news habit. Query that crowd. Do not ask one model to play the panel.

  • 連接

    託管版：加一個 URL，不必安裝任何東西

    URL
    https://api.crowdos.ai/mcp

    MCP 端點由 API 本身提供，因此永遠與平台同步——沒有需要更新的套件。在 claude.ai（設定 → 連接器）、Cursor，或任何支援遠端伺服器的用戶端中，把它加入為遠端 MCP 伺服器。支援 OAuth 的用戶端會在首次連線時引導你完成登入與授權；取得的授權會以可撤銷的金鑰形式顯示在 /account/api-keys 下。改以標頭傳送的用戶端，則使用 Authorization: Bearer crowd_... 搭配 MCP 金鑰。

  • 安裝

    自行架設：一行 pip，一段設定

    BASH
    pip install crowdos-mcp

    接著在 /account/api-keys 建立一組 MCP 金鑰（所有方案皆可建立），並把下面這段貼到你的 Claude Desktop 設定檔（~/Library/Application Support/Claude/claude_desktop_config.json）中：

    JSON
    {
  "mcpServers": {
    "crowdos": {
      "command": "crowdos-mcp",
      "env": {
        "CROWDOS_API_KEY": "crowd_..."
      }
    }
  }
}

    Cursor 與 Cline 在各自的 MCP 伺服器設定中使用相同的 JSON 結構。重新啟動宿主程式，CrowdOS 的工具就會出現在代理的工具選擇器裡。

BY RESEARCH JOB

The same tools, grouped by the question

The list below is every tool the server exposes. These four groups are the research jobs an agent is usually hired for.

  • OPINION

    What do people think?

    run_focus_group is a voting study: a sentiment summary and representative quotes. run_debate is a multi-round debate: convergence and the shift from round to round.

  • CHOICE

    Which option wins?

    compare_options is an A/B/C/D test on text options: a Borda leaderboard and a winner. rank_items is MaxDiff: a rank order with bootstrap rank intervals.

  • PRICING

    What would they pay?

    test_pricing runs a Van Westendorp or Gabor-Granger study: price points and a revenue curve.

  • BUDGET

    What will this study cost?

    preview_cost estimates the cost before the agent commits. It is free to call.

工具

15 個工具，為代理而設計

  • run_focus_group所有方案

    情緒傾向 / 投票調查——sentiment_summary 與代表性引述

  • run_debate所有方案

    多輪模擬辯論——收斂度與各輪立場變化

  • compare_options所有方案

    對 2–10 個文字選項（方案上限：Free 3 個，Pro 5 個，Max 10 個）做 A/B/C/D 測試——Borda 排行榜與勝出選項

  • test_pricingPRO + MAX

    Van Westendorp / Gabor-Granger 定價調查——價格點與營收曲線

  • rank_itemsPRO + MAX

    對 4–200 個項目做 MaxDiff——排序結果附拔靴法排名區間

  • pretest_questions企業版

    預先測試 1–40 題問卷草稿——正式發放前給出逐題結論

  • estimate_reach所有方案

    預覽 custom_audience 篩選條件所描述的受眾有多寬——免費

  • preview_cost所有方案

    正式送出前估算費用與耗時——代理在這裡編預算

  • list_demographic_presets所有方案

    查看可用的受眾範本

  • get_simulation所有方案

    取得先前已執行調查的完整結果

  • cancel_simulation所有方案

    停止執行中的調查並釋放保留額度

  • crowd_sample所有方案

    瀏覽即時的代理人群（已去識別化）

  • list_panels企業版

    已儲存的面板（由 persist_panel / evolve_panel 執行建立）— 僅中繼資料，最新在前

  • get_panel企業版

    單一已儲存面板 — 規模、到期時間、人口組成、已記錄的工作階段（不含受訪者名單）

  • delete_panel企業版

    刪除你擁有的已儲存面板 — 立即且永久

preview_cost 是多數代理最少用到的一個——在任何樣本數超過 100 的調查之前先呼叫它，看清楚費用（美分）與耗時（秒）的估計值，再判斷預算是否允許。呼叫免費。

實際運作

一次代理執行長什麼樣子

TEXT
> 就企業是否應強制實施週休三日執行一次焦點團體。
   使用 us_general_population 裡的 200 個代理。

[tool: preview_cost]   → estimate_cents=400 · ~33-88s
[tool: run_focus_group] → sim ad4b3736…
   sentiment_summary: { positive_pct: 71.5, negative_pct: 18.0,
                        neutral_pct: 10.5 }
   sample_responses: [ Maria Chen "It would be great for ...", ... ]
   billing.actual_cents: 400

代理會先選擇 preview_cost，因為樣本數並不算小；確認 4.00 美元在可負擔範圍內之後，才送出 run_focus_group。精簡過的回應（情緒摘要 + 5 則代表性引述 + 計費資訊）足以放進脈絡視窗；若代理需要每一位受訪者的完整推理，可用 get_simulation 取回完整資料。

ACCURACY

The opinion benchmark, published

On the Pew opinion benchmark, the crowd's opinion distribution matches published polling at 93.3% parity, with a mean absolute error of 3.36 percentage points and a Pearson r of 0.984. That published figure is the opinion benchmark. Creative ranking and in-market ad performance are measured separately.

費用

所有方案皆可使用

MCP 適用於所有方案，包括 Free。調查會以與網頁版相同的每位受訪者費率從帳戶餘額扣抵——一場小型焦點團體約 0.40 美元。

樣本組上限與網頁版一樣依方案而定：Free 最多 20 位受訪者，Pro 500 位，Max 1,000 位；每分鐘請求數分別為 5、30、120 次，每月 MCP 消費上限分別為 $5、$500、$2,500。價格 →

每一次 run_focus_group / run_debate 的回應都包含 billing.actual_cents 欄位，方便代理自行累計花費。

框架

REST API 僅供企業版帳戶使用

想從自己的程式碼而不是代理宿主程式呼叫 CrowdOS？REST API 開放給企業版帳戶使用。企業版 → LangGraph 與 CrewAI 都有社群維護的 MCP 橋接套件，不必執行 MCP 宿主程式，就能把我們的伺服器掛進框架代理裡。

WHERE THIS FITS

The same crowd, in the browser

The MCP server is the way an agent runs a focus group. Pulse is a vote, a debate or a comparison. Screening Room pressure-tests a creative, a concept or a deck before you ship. Ads Labo reads which of two to four ad variants a panel prefers.

FAQ

What is a focus group MCP server?

A focus group MCP server is a Model Context Protocol endpoint an AI agent can call. CrowdOS speaks it two ways: a hosted endpoint the API serves directly, and a pip-installable server for desktop hosts. Both offer the same tools for your account. Claude Desktop, Cursor, Cline, Zed, LangGraph, CrewAI, AutoGPT and Devin speak the protocol. Drop the server into the host's config and its tools show up in the model's tool picker.

Does it work with Claude, Cursor and Cline?

Yes. Claude Desktop, Cursor and Cline speak Model Context Protocol natively. Cursor and Cline use the same JSON shape in their MCP-server settings. Restart the host and the CrowdOS tools appear in the agent's tool picker.

How is CrowdOS different from LangGraph or CrewAI?

CrowdOS is the crowd an agent should query. LangGraph and CrewAI are hosts: the place the agent runs. LangGraph and CrewAI have community MCP-bridge packages that let you mount the CrowdOS server into a framework agent without running an MCP host.

Why not prompt an agent to role-play a panel?

One model acting like N personas (effectively n=1). An LLM persona is a costume for one prompt. Ours is a database row: census-grounded people with a fixed identity, an opinion history, and a standing news habit. Query that crowd. Do not ask one model to play the panel.

How accurate is the crowd an agent queries?

On the Pew opinion benchmark, the crowd's opinion distribution matches published polling at 93.3% parity, with a mean absolute error of 3.36 percentage points and a Pearson r of 0.984. That is the published opinion benchmark. The full ledger is on the benchmarks page. Creative ranking and in-market ad performance are measured separately.

30 秒拿到你的 MCP 金鑰。

免費註冊 →已登入？管理金鑰 →