Web-agent vendor TinyFish publishes an open, 300-run Mind2Web benchmark score; its page names no compliance certification or insurance terms

TinyFish builds infrastructure for AI agents that navigate and act on live websites on a company's behalf. "Navigate pages, fill forms, authenticate into sites and return structured results," the company's site states, describing an agent "built to complete production tasks accurately at scale."

The company backs its accuracy claim with a specific benchmark citation and an offer to check it. TinyFish says its agent scores "89.9% overall on Mind2Web, the industry-standard benchmark for web agents," and "81.9% on hard tasks, where Operator scores 43.2% and Claude scores 32.4%." The company adds: "All 300 runs are open-sourced and independently verifiable" — a transparency practice AI Dispatch has found in few agent vendors making benchmark claims of this kind.

TinyFish supports the Model Context Protocol, saying it is "MCP-native, so it works with Claude, Cursor, and any MCP-compatible client," and separately offers a "Direct API available for custom integrations." Its site states "Search and Fetch APIs · Now free," while its Agent and Browser products remain live but credit-metered. Pricing starts with "500 free credits to start," after which the company says users "Pay credits only when you run agents or browse with sessions" — free search and page-fetch functions are not themselves metered.

The company's site names no compliance certification, no audit-trail mechanism, and no insurance carrier, coverage limit or indemnification term, according to AI Dispatch's review of its published material — absences scored under AI Dispatch's compliance and insurance-indemnity dimensions, not partial credit toward them.

TinyFish delivers the agent as a hosted service: its site names "Cloud browser sessions" among its offerings, with no self-hosted or on-premises deployment option stated on the page reviewed.