Dropper

I gave an AI agent 2,000 credits and 90 real API call on @Nevermined_ai. It spent 708 credits, about 71 cents. Disclosure: Nevermined provided the credits for testing. 2,000 credits is about $2 at their own pack price ($10 for 10,000). No card, no wallet, no top-ups. Everything below ran the way an agent would call it, through their REST API and MCP server. What I tested Easy: IP lookups, a weather check, geocoding. Typical: a research agent across three search engines plus a scraper, a due diligence check on a real website, an SEC filing search, a DeepL translation, AI images, and auto-pick, including asks it can't do. Hard: bad requests, wrong paths, oversized requests, the same request sent twice, an LLM call, a fake API key, a full MCP run, a one-prompt "US Creator Weekend Brief", a 40-call reliability loop, and a full ledger audit. What it does well ➢ Billing is exact. Every charge in the ledger added up to my balance drop, to the credit. Every quote matched the charge. ➢ Reliability was near perfect. 88 of 89 valid paid calls came back. The one failure was refunded automatically and the retry worked. The reliability loop went 40 of 40. ➢ Failed calls are mostly fair. A bad request was refunded. Wrong paths, missing fields and oversized requests were refused before any charge. ➢ Duplicate protection works. Same request twice, one charge. ➢ The one-prompt test is the real win. Two search engines, a scraper, an LLM, DeepL, an image and file hosting: 8 calls, 93 credits, 79 seconds, zero vendor signups. It found 3 real upcoming US creator events, with a Spanish hook and a hero image. ➢ It's built for agents. The MCP server tells the agent its balance, the cost and when credits expire, in plain sentences. ➢ It says no honestly. "Book a hotel room" came back as no match, charged nothing. What it does badly ➢ Slow. Every paid call takes about 4 seconds. The same IP lookup straight from the vendor took 0.2 seconds. ➢ Headline prices are off. DeepL's card says half a cent. I paid 5 cents, 10x the headline. ➢ The markup is hidden. On average I paid 1.68x what vendors received. The LLM call was about 24x. ➢ You pay for junk if the vendor says 200. A paywalled scrape, two useless SEC searches and a geocode that put a well known neighborhood in the wrong country were all charged. The brief's LLM also got one event date wrong. ➢ "No stablecoins" covers part of the shelf. Credits pay 135 of 160 services and 61% of endpoints. ➢ No spending cap on credits. A looping agent could drain the whole balance. ➢ Dashboard rough edges. It says "No purchases yet" while Purchases lists every call, and "Total spent" shows a blank dash. What 1,000 credits (about $1) buys at measured prices ➢ 500 IP lookups ➢ 333 page scrapes ➢ 200 AI images ➢ 125 short GPT-4o-mini answers ➢ 100 Exa searches ➢ 20 DeepL translations ➢ 12 website screenshots ➢ About 10 full US creator briefs Scores (my judgment) Onboarding 8 Discovery 8 Agent experience 8 Billing accuracy 10 Reliability 9 Failed-call fairness 7 Price transparency 5 Latency 4 Coverage on free credits 6 Budget controls 4 Docs accuracy 6 Verdict: 7/10 The plumbing is the best I've tested. Exact money, almost nothing broke, and one key replaced 13 vendor signups. It loses points on speed, headline prices that don't match what you pay, and charging for useless "successful" answers. ➢ Use it if you're building an agent that needs many data sources without signing up for each one. ➢ Skip it if you need speed, heavy use of one vendor, or a hard spending cap without a crypto wallet. What I'd fix first Show the real per-call price on every card Add spending caps for credits, not just USDC Cut the 4 second overhead per call Pass vendor error messages back to the agent Fix catalog paths the router refuses Let agents flag junk 200 answers for a refund Would you let your agent spend money on its own? Tell me below 👇