Insight, not access — why I built a paid MCP for eBay sellers.

One solo developer, one eBay store, and the small handful of decisions that shaped a paid connector for Claude.

The Nemadji vase had been sitting in my store for over a month. Nice piece — blue-green swirl, Minnesota USA, matte finish. Not moving. I had two more just like it. Also not moving. I was in Claude asking questions about my own store, and I typed:

"Which items should I bundle?"

Claude gave me a thoughtful, generic answer. It didn't know I had three Nemadji vases sitting there. It didn't know they'd been listed for a month. It didn't know my store existed.

That was the moment I decided to build Flip Intel.

The problem was old. The shape was new.

I've been selling on eBay part-time for a while. I know the feeling of staring at a spreadsheet trying to answer questions like what actually sold last month? or which listings have views but no sales? or what's my fee rate on this category, really?

The tools that exist are fine. eBay's Seller Hub is a decent dashboard. Terapeak tells you what other people are selling. If you're a real business — the kind with an accountant and a shopping-cart plugin — CData will let a developer wire your eBay data into a warehouse for $499 a year. Improvado does it for you if you get on a sales call first.

None of that works for the person I'm building for: the side-hustle seller doing $2,000–$10,000 a month, part-time, on the couch after work. That person doesn't run a warehouse. That person doesn't want to log into a dashboard at all. That person wants to ask a question the way they'd ask a coworker — "which listings are dead weight?" — and get a real answer about their actual store.

MCP made that shape possible for the first time. I could build the tool where the seller already spends their time — inside Claude — instead of asking them to open one more tab.

325 tools, zero revenue

Before I started, I looked around. There was already an open-source eBay MCP on GitHub. 325 tools. Zero revenue. Impressive in scope, impossible to onboard. If you're not a developer, you can't run it. If you are, you're pasting a config file into your terminal, spinning up a local process, hoping the OAuth doesn't drift.

That's the shape of most MCP servers right now — impressive but developer-only. The insight I couldn't unsee was:

So I built one. But I anchored on something different than the raw-wrapper approach.

Insight, not access

Access isn't a business. Access is a feature. What sellers want is answers. So the first tools I built were shaped like answers:

• sales_summary(period) — not "give me all orders" but "here's what you sold, with context and a comparison to last period."
• top_products(period, metric) — ranked, with a "so what."
• slow_moving_inventory — the question every seller asks, phrased the way sellers ask it.
• fee_anomalies — the one that surfaces problems before you notice.

When Pro-tier landed, the shape stayed the same. suggest_bundles finds slow movers that fit together. generate_bundle_collage produces the composite photo. create_bundle_listing publishes the result. Three tools, one workflow, one conversation. On June 8th, I published my first Pro-tier bundle — three Nemadji vases into one live listing — end to end in a single Claude session. That's the product.

The pricing bet

The hardest early decision was the price. Every MCP that charges money is guessing right now — there's no comp for "one Claude connector that runs your store." So I looked at what my ICP already pays: $20 a month for Claude Pro. I decided Flip Intel had to be cheaper than that at entry, so it reads as an accessory to Claude and not a second subscription decision.

Starter is $9. Growth is $29. Pro is $79. Everyone gets the same first seven days for free, with a card on file.

No free tier. Not because I don't like free — because the free-tier plan attracts people who won't ever pay, and the free-tier customer costs the same to serve as the paying one. The seven-day trial does two things I couldn't get otherwise: it pre-qualifies (if you'll put a card down, you'll probably keep the subscription), and it gives me a real deadline to prove value. Every seller who signs up has one week to have a moment where the tool did something real for them. If that moment doesn't happen, the tool isn't good enough yet — and I'd rather learn that from a paying trial than from a free-forever ghost.

Billing is a product decision

I made a decision most bootstrapped SaaS makers don't: I don't do usage-based billing. Every tier gets a flat monthly rate and the full tool surface for that tier. If your bundle listing gets you $200 in a month, you paid me $9 or $29 or $79 that month — same as the seller who barely opened Claude.

The reason is philosophical, but there's a business version. Metered billing means I have to build a billing pipeline, a rate-limiting layer, an overage notification system, and a customer-support playbook for disputed usage. Flat billing means I have to build tools that make the flat price feel like a bargain. I'd rather do the second thing.

Billing still has a shape, though. When a card fails, dunning is soft: 5 days, 3 retries, transparent emails. When a subscription is fully canceled, I don't nuke the account — I preserve the eBay connection and the seller's history for 90 days. A card update reactivates access in under a minute. The seller who canceled six weeks ago and then changed their mind gets their store back, not a tombstone.

Use the cheapest infrastructure that works

I run everything on Fly.io for about $15 a month. Postgres for the database. Cloudflare Pages for the marketing site. Two Cloudflare Workers cron jobs for lifecycle sweeps and trial-email sends. R2 for the composite photos the Pro bundle workflow generates.

Fly.io because it's cheap and "always-on" is worth more than fancy autoscaling when you have ten customers. Cloudflare because R2 doesn't charge egress, and I don't want to be surprised by an eBay-CDN-shaped bill.

If you're building a paid MCP right now, this is the thing I wish someone had told me:

The temptation to reach for Kubernetes, feature flags, and a full observability stack is real. Resist it until you have a hundred paying customers. Boring infrastructure is the friend of the bootstrapped founder.

The moments that shaped the tool

A lot of the product exists because I broke things on my own store first.

The first time I tried to publish a bundle via the API, eBay rejected the shipping packageType. I had MAILING_BOX set. The seller's fulfillment policy was USPSParcel, which — as I learned by reading the error five times — does not accept MAILING_BOX. Now the tool auto-picks a packageType that the seller's actual policy will accept, before the request ever reaches eBay.

The first version of suggest_bundles picked an eBay category for the Nemadji vases that was under Costumes, Reenactment, Theater. Costume category. For pottery. That's an agent using the wrong tool. So I built suggest_ebay_category — a tool that verifies category fit against eBay's taxonomy before listing anything.

The first live bundle I shipped went out unpublished — the offer was created on eBay's side, but the publish call failed. I had to write a one-off recovery script that lists the stuck offer state, prints the plan, and waits for a yes/no before it either publishes or aborts. That "list first, confirm, then act" pattern is now the shape of every destructive tool I build.

Every one of those was a bug in my own store first. Then it was a fix. Then it was a small chunk of shared infrastructure other sellers wouldn't have to hit.

What honest looks like

Two things I want to say plainly, because most founder posts don't.

First: the price point for Pro at $79 is a guess. I don't have enough Pro customers yet to know if it should be $59, $79, or $99. Every conversation I have with a Pro seller updates my belief. If you're reading this and you have a strong opinion — either that it's too high for what's there today, or that I'm underselling it — that's the kind of email I want.

Second: this is a solo project. I built the product, I run the ops, I answer the support email, I write the marketing site. There is no team behind Flip Intel. That means the response time to your bug report is measured in hours, not weeks. It also means if you sign up on a day I'm sick, the roadmap that day is slower. I'd rather be plain about that than pretend otherwise.

What's next

Right now I have 15 tools live across three tiers and a growing set of paying customers. The next three things shipping, in order:

• Proactive alerts — a weekly digest email that fires automatically when your revenue drops or your fee rate spikes, so sellers who don't open Claude every day still catch the signal.
• Stored revenue history — so you can compare this December to last December without hitting eBay's API every time. Right now every query is live; stored snapshots make trend questions instant.
• Sourcing brief — a Saturday-morning tool that synthesizes your last 90 days into a plain-English buying guide: which categories to prioritize at the estate sale, which to skip, what acquisition price range keeps your margin intact.

The MCP ecosystem is still finding its shape. Whether "paid MCP servers" become an ordinary business or an odd historical footnote depends on whether the products built on top of MCP are better than the raw APIs they wrap. I'm betting they are — because insight beats access, hosted beats DIY, and answering the seller's actual question beats surfacing 325 tools.

If you sell on eBay part-time and any of this resonated, the trial is seven days with a card on file. The first bundle the agent helps you ship usually pays for the year.