How to enable your agents to talk to real people
Your agent can reason over data, search the web, call APIs, and write you a report. The one thing it usually cannot do is ask a real person what they think. So the moment a question needs human judgment — Would anyone actually buy this? Which message lands? What would make them switch? — the loop breaks, and you leave the conversation to set up a study by hand.
Flashpoint.AI closes that gap. It gives Claude a set of tools to design a study, recruit a real audience, field it, and bring the results back — all inside the same conversation. No API wiring, no dashboards, no copying tokens between tabs. You describe who you want to hear from, and your agent goes and asks them.
Here is the full loop, start to finish.
1. Connect Claude to Flashpoint.AI
Flashpoint.AI runs a hosted Model Context Protocol (MCP) server, so this is a one-time setup.
Claude Code: run the command below, then type /mcp to authenticate. You should see Authentication successful. Connected to flashpoint.
claude mcp add --transport http flashpoint https://flashpoint.ai/mcpClaude Desktop: open Settings → Connectors, click Add custom connector, and paste the endpoint below.
https://flashpoint.ai/mcpYour session refreshes automatically after the first sign-in. If a tool ever reports that it is unauthorized, reconnect and sign in again. Claude Code and Claude Desktop are supported today; the Claude.ai web connector and ChatGPT custom connectors are not yet supported.
2. Describe the study — let the agent build it
Once connected, you do not fill out a form. You say what you want:
"Create a 5-question concept test for a new ready-to-drink cold brew, targeting US coffee drinkers."
Claude drafts the full instrument, validates it, and saves it as a draft in your workspace. The key detail is that it does not just map your request into five text boxes — it applies survey design logic a good researcher would use.
For this study it produced:
- A multi-select on which coffee formats respondents currently buy (ground/whole bean, single-serve pods, ready-to-drink cans, instant, cafe purchases)
- A single-select on what matters most when choosing a coffee product (taste, price, caffeine strength, convenience, ingredients, brand)
- A single-select on how often respondents buy ready-to-drink coffee
- A 5-point purchase-intent scale: how likely respondents are to try a new premium cold brew at $3.99
- An open-ended question: what would make respondents switch coffee brands
Along the way it randomized the format and attribute lists to reduce order bias, added None of these and write-in options, used a balanced 5-point scale for purchase intent, and placed the open-ended question last to minimize drop-off.
It can also attach a category screener — recent coffee buyers only — so you do not pay for respondents outside your market. The draft passes validation and stays unpublished until you decide to launch.
3. Quote the audience before you spend a cent
When you are ready to recruit, ask Claude to find respondents and recommend filters. For this study the agent proposed:
- Country of residence: United States
- Age 21-54: core category buyers
- Category screener: bought coffee in the past month
Demographics beyond that — sex, education, region — are left open on purpose. Broad filters fill faster and keep cost down, while the screener still guarantees each respondent is in-market. You can tighten later by region or buyer frequency.
Then Claude returns a transparent quote before anything is charged:
- 200 completes
- Approximately 62,000 eligible panelists match the filters
- Approximately 5 minute interview length
- $1.57 per complete -> $314.00 total
Pricing is linear, so you can soft-launch a 25-complete pilot for about $39 to pressure-test the instrument first, or scale to 1,000 completes for roughly $1,570.
Nothing is billed at this stage — the quote is only a quote.
4. Approve through secure checkout
Say the word and Claude generates a payment link:
"Pay $314.00 and launch your panel — link expires in 1 hour, and recruiting begins automatically once payment clears."
Once you have paid, a simple done brings the conversation back to life:
"Payment confirmed and your panel is live. Completes: 0 of 200. Responses should start arriving shortly given the 62K+ eligible pool."
5. Watch it field — then turn answers into decisions
From here, the conversation becomes your live dashboard. Ask for a count any time, and as completes land you can have Claude pull responses, run frequency breakdowns on closed-ended questions, cross-tab by segment, and theme open-ended answers into clear takeaways:
"We are at 140 of 200. Taste and convenience are outweighing price, purchase intent is strongest among current ready-to-drink buyers, and the open-ends repeatedly flag lower sugar and cleaner ingredients. Want a cut by purchase frequency or current format?"
That is what makes this more than a survey tool: the same agent that designed the study reads the answers and recommends what to do next.
The whole loop, in one place
Connect once. Describe the study. Review the filters and the quote. Approve through secure checkout. Then let your agent field the study and turn real human feedback into your next decision — without leaving the chat.
When the agent itself is the buyer
Everything above assumes a human driving Claude. But the agent can buy research on its own — a long-running process with its own wallet and no human in the loop. Skip the connector. The payment path is different, and simpler.
This already works with real money. On 2026-06-05 an agent paid us $0.05 USDC for a demographics call — settled on-chain, on Base mainnet, with no human involved. Same MCP server, same tools as the walkthrough above. Only the payment rail changes.
Agents find us without being told to. Flashpoint.AI is listed in Coinbase Developer Platform's Bazaar — the x402 catalog agents query to discover paid services on Base mainnet — and on agentic.market. Any agent running the Coinbase Agentic Wallet Skills finds us by tag (research, surveys, demographics, panels) and calls our tools directly. No account. No API key. No claude mcp add.
Each call costs USDC on Base mainnet:
chat— $0.01 per call. Orchestrates demographics, panels, and presentations in one request.demographics— $0.05 per call. Worldwide population, age, income, and education, with AI insights.presentation— $0.25 per call. Native PowerPoint decks from a data context.tool_survey— $0.50 per call. Designs, launches, distributes, analyzes, and exports a full survey.panel— quoted per request. Scales with sample size and provider, on the same $1.57 × n math the human flow uses above.
Settlement is the standard x402 handshake. The agent sends an unauthenticated request; we answer with 402 Payment Required and on-chain settlement terms; the agent signs an EIP-3009 transferWithAuthorization from its wallet and retries with an X-PAYMENT header; the Coinbase facilitator settles to our payout address on Base mainnet. No bridge, no escrow, no multi-step confirmation. One signed transaction per call.
Through the Claude connector, your Flashpoint.AI subscription covers usage — nothing extra per call. On the x402 path, agents pay per call in USDC, as listed above. See the Quickstart for setup, supported tools, question types, sample sources, and exports.