Rebuild Carvana's Deal-Finding Machine With an AI Agent and a Marketplace MCP

You can copy the mechanical core of a used-car business, sourcing cars below market value, using a single AI agent wired to a Facebook Marketplace tool. The video frames Carvana as a company that buys cars, cleans them, and resells them for a spread the creator pegs at roughly seven grand per car. Its pitch: automate the sourcing and arbitrage step yourself with an agent that scans listings every morning, checks each one against Kelley Blue Book, and pings you only when it finds a real deal. The build shown is a chain of small, scheduled "skills" running on top of an existing open-source Marketplace connector.

The Carvana Loop It Is Copying

The opening frames show Carvana's glass car vending machine tower with the caption "vibe code a $20B vending machine." The creator breaks the business into a visible loop: buy the car, collect it, clean it, maybe fix something, then sell it for more. Carvana is a real online used-car retailer known for those multi-story vending machines, where buyers collect a purchased car and can drop off a trade-in. Its money comes from reselling used cars for more than it paid, plus financing. The video's argument is narrow and worth stating plainly: you are not rebuilding the logistics, the reconditioning, or the vending machines. You are copying only the part where someone spots an underpriced car before anyone else does.

The One Prompt That Runs the Whole System

The clearest frame is a card titled "PROMPT 01 · THE WHOLE THING" with a Copy button and the instruction "Paste into your agent. Swap the {slots}." The six numbered steps are the actual spec:

Every morning check Facebook Marketplace for these searches: {my car list}.
Save every new listing (price, mileage, year, trim, photos, location, seller) to a database.
For each new car, look up its Kelley Blue Book private party value and flag anything more than {15%} under.
For anything flagged, message the seller, ask if it is still available and why they are selling, and negotiate toward {my max}.
Text me only when you need a decision or the seller says yes.
Save this as skills and run them on a schedule.

The bracketed slots are the only things you customize: your target cars, your discount threshold, and your walk-away price. Everything else is fixed logic. This matches the transcript, which describes the agent taking your keywords, searching Marketplace, checking Blue Book, messaging sellers, and texting you when it finds a deal.

The Connector Doing the Marketplace Work

A frame lists four GitHub repositories the agent can plug into, presented as real off-the-shelf options rather than something you build. The first, jdcodes1/facebook-marketplace-mcp, is labeled "graphql, no browser at runtime." That is accurate: it talks to Facebook's internal GraphQL API using your existing session cookies instead of driving a browser, and it can save a search as a monitor to track new listings over time, which is exactly what step one needs. The other repos shown are jlsookiki/secondhand-mcp (Marketplace plus eBay and Depop), hagaihen/facebook-mcp-server (Graph API, pages), and gomarble-ai/facebook-ads-mcp-server (Meta ads data). The point of the frame is that the hardest integration, reading Marketplace programmatically, is already solved by someone else. Note that these tools work by using your logged-in Facebook session, so the agent is acting as you.

Breaking It Into Scheduled Skills

Two later frames show the system as a grid and a ring of small named units: watch-marketplace (every morning), save-listing (price, miles, trim), vs-kbb (free values), flag-undervalued (15 percent under), message-seller, negotiate (toward my max), and a price-drop tracker for slow movers. The transcript's advice is to "turn each of these steps into individual skills, then schedule them to just run autonomously." This is the operational idea underneath the demo. Instead of one long prompt you babysit, each step becomes a discrete, reusable skill, and a scheduler fires the chain on its own each morning. The payoff frame is a phone notification titled "Found one." reading "2019 Tacoma, $4,20... Want me to offer $2..." The agent only interrupts you at the two moments that need a human: a decision, or a seller saying yes.

Where the Value Check Actually Happens

The flagging logic hinges on one comparison. A frame shows a Facebook Marketplace search for "bentley" with several listings, then the video cuts to kbb.com and the caption "Use its browser tools to go to kellybluebook.com." The agent pulls each listing's details, looks up the car's Kelley Blue Book value, and keeps only cars priced meaningfully below it. The prompt specifies private party value, which is the KBB category meant for a sale between two individuals and is the correct benchmark for a Marketplace listing rather than dealer or trade-in pricing. The 15 percent threshold is the single lever that decides how aggressive the agent is: raise it and you get fewer, better flags; lower it and you get more noise.

Key Takeaways

The system copies only Carvana's arbitrage step, finding cars priced below market, not its reconditioning or logistics.
One prompt with three variables ({car list}, {discount %}, {max price}) defines the entire agent; everything else is fixed logic.
An existing open-source MCP connector handles the hard part of reading Facebook Marketplace, so you do not build the integration.
Kelley Blue Book private party value is the benchmark, and a 15 percent-under threshold is the flag.
Splitting the flow into scheduled "skills" lets it run autonomously and only text you for a decision or a confirmed sale.
Because the Marketplace tools use your own Facebook session, the agent messages sellers as you, which carries real account and etiquette risk.