# CrowdOS > CrowdOS is creative pre-flight and A/B testing that an AI agent runs > mid-task. Over the CrowdOS MCP server (https://api.crowdos.ai/mcp, or > `pip install crowdos-mcp`), on every plan including Free, an agent tests > ad copy, messaging, concepts, and image and video ads against a panel of > census-grounded synthetic respondents who disagree with each other — not > one model's opinion — and gets back the totals and a ranked winner. Ads > Labo (test_ads over MCP) returns a ship, tied or rework call between two > to four variants, each variant's share of the picks and a head-to-head > test of the top two. CrowdOS reports relative preference, not in-market > CTR, CPM or ROAS. On the Pew opinion benchmark (an opinion measure, not a > creative-ranking measure), the crowd's opinion distribution matches > published polling at 93.3% parity (3.36pp mean absolute error, > Pearson r 0.984). ## Pricing - Free accounts get 20 credits every week, reset Monday 00:00 UTC — no card. Pro adds 1,350 credits a month, Max 5,500. Studies use credits, shown before launch: https://crowdos.ai/pricing ## Pages by question - Can my AI agent run a focus group? https://crowdos.ai/mcp - How do I put a focus group inside an agent's task? https://crowdos.ai/guides/ai-agent-focus-group-mcp - Guides index: https://crowdos.ai/guides - What is a synthetic focus group? https://crowdos.ai/synthetic-focus-groups - Which AI focus group tools exist? https://crowdos.ai/ai-focus-group-tools - Which AI tools do public opinion research? https://crowdos.ai/public-opinion-research - How do I test creative and messaging before launch? https://crowdos.ai/pre-launch-testing - How do I test ad creative (copy, image and video) before launch, over MCP on every plan? https://crowdos.ai/ad-creative-testing - Can I A/B test ad creatives before spending on traffic? https://crowdos.ai/ab-testing-without-traffic - How does a solo founder's AI agent test ad copy, images and video over MCP? https://crowdos.ai/solo-founders - How do I validate demand before I build? https://crowdos.ai/market-validation - What are the alternatives for synthetic research? https://crowdos.ai/alternatives - CrowdOS vs Synthetic Users: https://crowdos.ai/compare/synthetic-users - CrowdOS vs Aaru: https://crowdos.ai/compare/aaru ## What it does - Pulse (https://crowdos.ai/pulse): opinion measurement — vote, debate, or compare across census-grounded respondents. From 20 credits a study. - Screening Room (https://crowdos.ai/screening-room): pressure-test any stimulus (creative, product mockup, policy draft, pitch deck) with a reasoned panel + expert lenses. From 44 credits a study. - Ads Labo (https://crowdos.ai/ads-labo): creative pre-flight for ads, in the web app and over MCP (test_ads). One to four image or video ads, with Meta-style targeting, go to one census-grounded panel that disagrees with itself; with two to four, the report says ship, tied or rework (with the panel's reasons in the web app). From 75 credits a test. - MCP server (https://crowdos.ai/mcp): a hosted endpoint (https://api.crowdos.ai/mcp), or `pip install crowdos-mcp`, gives AI agents (Claude, Cursor, etc.) tools over the Model Context Protocol to A/B test copy (compare_options), read a concept (run_focus_group, run_debate), screen images and video (screen_creatives), pre-flight ads (test_ads) and, on Pro and Max, run MaxDiff (rank_items) and pricing tests (test_pricing). Every plan — Free, Pro and Max — can use it, at that plan's web limits. Results are totals only (counts and shares, the leaderboard and winner, a debate's synthesis), without individual answers. - REST API (https://crowdos.ai/enterprise): the same crowd over REST with bearer-token auth, SSE streaming, and webhooks, for Enterprise accounts. - Customer digital twins (https://crowdos.ai/solutions/customer-digital-twins): cohort twins built from a business's own customer file, one persistent archetype per weighted cohort (never one per customer), to rank offers and find objections before a campaign. In build with design partners. The same page describes how they are delivered: Digital Twins powered by CrowdOS, a private twin model built for one business, in Lite, Standard and Private tiers. Pilots by application. - Atelier by CrowdOS (https://crowdos.ai/atelier): the bespoke services arm that delivers customer digital twins, its first solution. ## Key facts - Studies run census-grounded synthetic focus groups of 20-1,000 AI respondents that answer in about two minutes. The Pew figures above are aggregate results on a 25-question battery. - A persistent backbone of 4,000+ archetypes reads real news and evolves on a recurring cycle (three times a week). - Respondents are sampled from published census distributions (US ACS, UK ONS, AU ABS and national equivalents across 24 markets) with validated Big Five (OCEAN) personality traits. - Every study draws a fresh, demographically weighted panel. In the web app, results include sentiment distributions, demographic crosstabs and verbatim per-respondent reasoning; over MCP, results are totals only, without individual answers. - The persistent crowd evolves only from exogenous signal — the news, life events, a peer feed, and time decay. It never learns from what customers ask it, so no study biases a later one. - Accuracy is published and reproducible, with the run id and date of the validating benchmark shown alongside the figure: https://crowdos.ai/benchmarks ## Docs - How it works: https://crowdos.ai/how-it-works - Methodology whitepaper: https://crowdos.ai/whitepaper - Benchmark ledger: https://crowdos.ai/benchmarks - MCP guide: https://crowdos.ai/mcp - Mid-task focus group (MCP how-to): https://crowdos.ai/guides/ai-agent-focus-group-mcp - Enterprise (REST API): https://crowdos.ai/enterprise - Customer digital twins: https://crowdos.ai/solutions/customer-digital-twins - Preparing a customer file for twins: https://crowdos.ai/docs/prepare-your-customer-file - Atelier by CrowdOS: https://crowdos.ai/atelier - The persistent population: https://crowdos.ai/the-crowd - Pricing: https://crowdos.ai/pricing