Cracked.ai

Cracked sits in a category that did not quite exist two years ago: the agent tool router. You do not pick Apify or Firecrawl or Serper — you write one call naming a capability and Cracked ranks the candidates by fit, price and measured health, then fails over if the chosen provider breaks. The measurement layer is the genuinely unusual part. Health is defined as the last 200 real runs of an endpoint, bucketed healthy / stable / degraded / outage / unknown with a median run time; the reliability index and the leaderboards are dated and link their sources. Ed25519-signed receipts on completed runs and a ledger where every debit carries the run id, units billed, provider share and fee mean an agent's spend is auditable, not a monthly surprise. Governance arrived in the September 2026 changelog: policies per workspace, key or customer tag with allowed and blocked capabilities, providers and tool ids, max cost per run, daily and monthly caps, allowed hours and approval thresholds; capability pins that force the smart-run router to a chosen provider; projects with a monthly budget cap that answer 402 PROJECT_CAP; and the policy decision contract at POST /v1/policy/evaluate that the runner consults before any credential is resolved. That is a supply-chain control story most routers do not tell. The cost model is where you have to be honest with yourself. Cracked bills the provider price plus 15% plus $0.001 per successful run, and its own worked examples show the gap: 1,000 Instagram profiles run $1.84 through Cracked against a $1.60 list price, and 100 Google searches run $0.50 against $0.10 at Serper's own rate. If you are already a direct customer of one provider at volume, routing is a service you are buying, not a discount. Bring-your-own-key with $0 markup is the fair version of that, but then you keep the provider's rate limits and support relationship rather than Cracked's master accounts. Waterfall runs are priced at the successful provider's cost plus 35%, or every attempt including misses, whichever is higher — cheap when it wins on the first try, less so when a capability is flaky. Second-order cautions: the catalog is young (first release September 1, 2026; 6 versions by September 2), so reliability numbers rest on thin history for many of the 170 capabilities, and a benchmark on September 2 found 136 of 159 capabilities measurable in 192 runs. Credit is spendable for 12 months and spent oldest-first, which quietly caps how much you should prepay. Where it fits: engineering teams shipping multi-capability agent workflows, agent builders on Claude Code, Cursor or MCP clients who want a working key in one call, and startups watching scraper and inference spend who value free failed runs. Where it does not: single-provider high-volume users, enterprise procurement that wants long operating history, buyers who need the underlying provider account itself, and anyone expecting a human dashboard product — this is built for agents calling an API.