Feature Description

A new Nexent agent does not get web search automatically. Today, someone has to enable a search tool or add a provider-specific skill before the agent can look something up. A native, keyless default would make that first search work without another provider signup.

Proposed Solution

Could Nexent add Parallel as a native web-search provider and use it as the default only for a new agent with no search provider selected? The anonymous Streamable HTTP MCP endpoint is https://search.parallel.ai/mcp; its web_search tool needs no Parallel account, API key, or auth header. This request is for search results only, not web_fetch.

Please keep Tavily, Exa, and Linkup available, along with user-added MCP tools, saved agent configs, and any explicit provider choice or search opt-out. The provider should fit the existing per-agent tool flow, appear in effective runtime configuration, and return useful results through Nexent’s normal search and citation handling. A server-listing example alone would not make a new agent able to search.

Each search sends its query, objective, and supplied context to Parallel’s hosted service, which returns search results to Nexent. The anonymous path is rate-limited, so this is not a promise of unlimited access.

Additional Information

The Artificial Analysis Search API snapshot updated 2026-09-22 reports these results. DeepSearchQA scores answer quality across multi-search research tasks. Time per Task estimates the whole research task, not one API call.

• DeepSearchQA average F1: Parallel Fast 0.8024847387112691 (measured 2026-08-19T07:47:39.059458Z), Exa Auto 0.7780452590902946 (2026-08-17T12:48:42.175571Z), Tavily Basic 0.7436967920018923 (2026-08-17T11:52:57.029511Z), and Linkup Standard 0.6892871432995258 (2026-09-17T03:31:44.550225Z). DeepSearchQA chart
• Time per Task: Parallel Fast 17.75260526326803 seconds (measured 2026-08-19T07:47:39.059458Z), Linkup Flash 24.67526756055607 seconds (2026-09-16T07:05:16.131565Z), and Tavily Basic 42.816239607412484 seconds (2026-08-17T11:52:57.029511Z). These are Search API benchmark results, not Nexent measurements. Time per Task chart

Proposed validation, not run in Nexent yet: add deterministic tests for the new-agent default and preserved existing choices, then exercise provider selection, effective configuration, the canonical web_search call, and useful linked results through Nexent’s actual caller using the keyless MCP path.

Btw - I'm a devrel engineer at Parallel. Would this default for new, unconfigured agents be a fit for Nexent?