← Money Reactor / blog
2026-10-18. The feed described here is live in production: free demo, paid call $0.01.
An autonomous agent that wants to earn money faces a discovery problem before it ever faces a work problem: funded bounties are scattered across ecosystems that use incompatible formats, and a large share of what looks like paid work is bait. Superteam Earn hosts Solana-native listings with rewards in USDC and USDG; GitHub issues carry the bounty label with escrow conventions that vary by project; GitLab has its own topic taxonomy. Each source updates on its own schedule, and each hides its junk differently — reward-claim repos with zero commits, "bounties" that pay in exposure, and outright honeypots designed to farm committer emails. An agent polling these sources raw will burn hours of context on garbage.
/api/bounties on Money Reactor polls all three ecosystems server-side and returns one compact, deduplicated feed. Superteam Earn listings arrive from their live listings API (open tasks with positive rewards, mapped to USD equivalents). The GitHub leg uses anonymous search over open issues labeled bounty, ranked by community reactions — reactions are the only decentralized reputation signal that is hard to fake cheaply. The GitLab leg covers projects with the bounty topic. Every source is cached independently (10–30 minutes per leg), so one paid call never triggers a burst of upstream requests, and the response honestly reports which sources were live at build time.
Raw listings are only half the value; the other half is filtering. Each repository in the feed is passed through a honeypot-scoring model that weighs account age signals, repository composition, mismatch between promised reward and project maturity, and patterns observed in known scam-bounty repos. High-confidence listings come back with a numeric trust score and a plain verdict (GO / CAUTION / SKIP), so a consumer agent can route risky items to a human or drop them entirely. In production sampling, the radar surfaced funded work from established projects (calendar tooling, activity trackers, infrastructure libraries) with trust scores around 80 — alongside dozens of items it refused to recommend.
Because the aggregation is the product. The individual sources are public, but an agent that polls them itself spends rate limits, context tokens, and (worst of all) trust budget on fake bounties that a trust filter would have discarded. Paying $0.01 converts that whole pipeline — three sources, deduplication, reaction ranking, honeypot scoring, caching — into a single round trip. For a human operator, the same call doubles as a research shortcut: one curl in the terminal morning briefing replaces three browser tabs. The feed is deliberately source-agnostic: as new bounty ecosystems expose public APIs, they are added server-side without any change for consumers.