Why build a search API?
Massive demand. No infrastructure.
Web search is the most fundamental agent capability — every framework ships it on day one. But the existing options are all compromised: Google CSE is expensive and rate-limited, Brave doesn't return Google results, SerpAPI doesn't own its infrastructure, Tavily summarizes away the raw data. We built the search API we wanted to use ourselves — Google-grade results, sub-500ms, 99%+ uptime — on infrastructure we own end-to-end.
Key Metrics
live arena dataLive Leaderboard
| # | Provider | Avg Latency | Win Rate | Success | Runs |
|---|---|---|---|---|---|
| 1 | TinyFish | 408ms | 75.3% | 94.8% | 77 |
| 2 | Firecrawl | 926ms | 10.4% | 98.5% | 67 |
| 3 | Exa | 1118ms | 0% | 100% | 67 |
| 4 | Tavily | 1147ms | 6% | 100% | 67 |
| 5 | Parallel | 3621ms | 0% | 100% | 67 |
| 6 | BrightData | 3476ms | 0% | 95.5% | 66 |
Latency Comparison
Cascade Fallback
Google and DDG are hit concurrently — whichever returns first wins. If both fail, falls back to vendor API.
Architecture
Browser Farm
Five warm Chromium instances with round-robin distribution and auto-recovery watchdog.
Why We Win
Warm pool eliminates cold start. Custom binary returns results as soon as DOM parses.
Concurrent engines + vendor API fallback + auto-recovery watchdog.
Own the browser, proxy, and infra. No per-query fees.
Design Principles
Competitive Comparison
| Provider | Approach | Quality | Cost Model | Reliability |
|---|---|---|---|---|
| TinyFish | Own browser + proxy infra | Google-grade | Own infra (no per-query fee) | Concurrent engines + vendor fallback |
| Google CSE | Official API | Native Google | $5/1000 queries | Rate limited |
| Brave Search | Own index | Good, not Google | $3/1000 queries | 99%+ |
| SerpAPI | Proxy scraping | Google-grade | $50/5000 queries | Proxy dependent |
| Tavily | AI-optimized | AI-summarized | $1/1000 queries | ~95% |
Watchdog & Auto-Recovery
Three layers of health monitoring ensure zero-downtime operation:
Proxy Infrastructure
ISP-grade proxies on datacenter — ISP reputation with datacenter speed.
Market Context
- Google CSE — expensive, rate limited
- Brave Search — good quality, limited control
- SerpAPI — proxy layer, not infra owner
- Own the browser — control latency end-to-end
- Own the proxy — control cost and IP quality
- Full stack — search + browser + agent in one
Deferred
Proprietary Crawling & Indexing
Own search index comes later when we understand query patterns and can justify the investment.
Edge Deployment
Edge compute for millisecond latency. Current single-region is sufficient. Requires validated demand.