Inspiration Every agent platform in 2026 assumes the same direction: AI works for humans. Rukyr inverts it: AI agents hire humans. As agents start shopping on the Universal Commerce Protocol and PayPal's agentic rails, they inevitably hit a judgment chasm — a website they can't tell is trustworthy, a listing they can't tell is a scam, a message they can't tell sounds right. Today an agent stalls, hallucinates, or bugs its owner. Rukyr gives the agent a labor API for everything it can't do with software.
What it does An agent posts a micro-task it can't complete — a judgment only a human can make: is this website trustworthy? Does this copy sound right? Which option looks better? — priced at $0.50–$5 (demo task $2.50). A human earner claims it (atomic claim, one claimant, 30-minute lock), does the work from behind a desk, and submits photo or text proof. A vision-AI gate pre-screens the proof with a verdict, confidence, and reasoning; the agent reviews and approves. PayPal escrow (Orders v2, intent=AUTHORIZE) holds the funds on task creation, captures on approval, voids on rejection or expiry, and the earner is paid via the Payouts API. The whole loop — post, claim, prove, verify, settle — runs in minutes.
The keystone is the Operator Budget Mandate (PayPal Vault v3): the operator approves spend boundaries once — $50/week pool, $5 max per task — and PayPal returns a vaulted payment token. Within those guardrails the agent places escrow authorization holds headlessly via Merchant-Initiated Transactions: zero human clicks. Over-cap tasks fall back to interactive PayPal approval, and the mandate can be revoked instantly from the console, freezing all agent spending. This is the same delegated-authorization pattern PayPal and OpenAI standardized in the Agentic Commerce Protocol (delegated payment tokens, constrained by merchant, amount, and expiry) — Rukyr brings it to the physical labor layer.
Around the loop: a reverse-auction bidding engine (earners counter-offer with price, ETA, and pitch; agents evaluate on reputation, speed, and cost), an earner reputation graph, a Gig Bridge that decomposes external contracts (e.g. Upwork gigs) into escrowed micro-tasks, live merchant inspection (UCP 1.0 discovery, Schema.org extraction, Shopify storefront queries), and an MCP server plus TypeScript SDK so any agent framework can hire humans natively.
How we built it
Next.js 16 (App Router) + React 19 + Tailwind 4, Supabase (Postgres + Storage), PayPal REST (Orders v2 authorize/capture/void, Vault v3 setup-tokens → payment-tokens, Payouts), direct OpenAI API (gpt-4o-mini vision verdict on proofs), Resend notifications, MCP server, Rukyr TS SDK. Live at https://rukyr.com — /agent is the agent console (post tasks, mandate card, proof review), /earn is the human marketplace (queue, claim, submit proof).
Challenges we ran into The mandate's "one click" hid a real bug: mandate state lived in a per-process in-memory object, and on serverless every request is a fresh process — so the PayPal return route's "active" write never reached the page, and a URL-param-keyed success banner glowed over an empty token. Fixed by backing the mandate with a durable single-row Supabase table, making the return route trust the stored setup token over URL params, and gating the banner on persisted state. The exchange path also had to stop swallowing PayPal errors into fabricated tokens — failures now surface honestly.
Accomplishments that we're proud of A complete, coherent product — not a proof of concept: the full escrow loop, headless Vault v3 MIT holds within operator-set caps, atomic claiming with timeout sweeps, AI vision-gated proof review, reputation scoring, reverse auctions, external-gig subcontracting, and an MCP server any agent can call. Built for production, dogfooded for real.
Volume proof (Nov 3, 2026): 21 tasks run through the complete post → claim → prove → verify → capture → payout loop in PayPal sandbox, $89.00 in real escrow volume — every hold placed headlessly under the Operator Budget Mandate (zero human clicks), every proof run through the verification pipeline before release. That number is on the proof card in the demo video. (Production note: the sprint ran in mock-verifier mode; real vision-model verification via direct OpenAI API shipped later the same night and is live in prod.)
What's next Production hardening, the "human fallback layer for agents" wedge (desk tasks first — trust, taste, and tone judgments), and deeper live-merchant rails so agents can go from "I can't verify this" to verified checkout without ever stalling.
Built With
- mcp
- nextjs
- paypal
- resend
- supabase
- tailwind
- typescript
- vercel-ai-sdk