September 2026 Summaries

WebMCP is an evolving browser API proposal that allows websites to expose deliberately defined, application-specific tools to AI agents, enabling tasks such as retrieving account usage, searching documentation, or updating page state without relying on manual clicks and form navigation. Unlike the Model Context Protocol (MCP), which uses independent local or remote servers and JSON-RPC to provide tools, resources, and prompts across environments, WebMCP operates within a live web page, can use its authenticated state and interface context, and does not automatically create an MCP endpoint. Websites can register tools through JavaScript functions or annotate HTML forms, while retaining responsibility for validation, authorization, permissions, confirmation flows, and protection against unsafe output or data exposure. WebMCP is best suited to collaborative, in-page workflows where the current application state matters, whereas MCP remains appropriate for services and automated tasks that must function independently of an open browser tab; organizations may use both approaches for different interactions. As of September 2026, WebMCP remains experimental, with limited browser availability, so developers are advised to feature-detect support, preserve standard interfaces, begin with focused read-only tasks, and evaluate reliability, security, latency, and user outcomes before broader deployment.

Company research APIs serve three distinct needs: discovering companies that meet complex criteria, looking up facts about known companies, and enriching existing company lists, with database providers and live web research APIs suited to different tasks. Databases such as People Data Labs, Crunchbase, Apollo, Coresignal, Diffbot, and ZoomInfo provide fast, structured, domain- or name-keyed records but may lack fresh or unusual information, while web research services including Parallel, Exa, Linkup, Tavily, Firecrawl, and Brave use repeated live searches and page retrieval to investigate current, multi-constraint questions at greater time and cost. Openbenchmarks’ August 2026 multi-turn company search evaluation found Parallel Search basic led search-only company discovery with a 46.5 F1 score and 88.7% precision, whereas Exa deep led when agents could fetch pages, scoring 48.2 F1, and Exa also led a separate company-news factual lookup benchmark at 99.2% accuracy. Results indicated that recall remains limited across providers, with the strongest configuration recovering roughly one-third of valid companies on average, underscoring the need for verification. The recommended approach is to use research APIs for discovery, recent news, and facts beyond fixed schemas, database APIs for inexpensive stable firmographic enrichment, and often both in production workflows, while testing tools against an organization’s own questions, cost, latency, and accuracy requirements.