[Feature]: Add Parallel Search MCP to Super Magic web search

[georgeatparallel]

Component

Super Magic

Feature Type

Integration with external service

Problem Description

At the pinned baseline, Super Magic's built-in web_search uses the configured engine, which falls back to magic. Parallel is not in the provider registry, so an administrator can't route the normal search tool to Parallel's free Search MCP today.

Proposed Solution

Please add Parallel Search MCP as an optional provider for web_search. The existing WEB_SEARCH_DEFAULT_ENGINE setting could select it and send the tool's queries to https://search.parallel.ai/mcp. Anonymous use doesn't need an API key. Please keep magic as the default and leave the other engine choices alone.

Use Cases

1. As a self-hosting administrator, I can set WEB_SEARCH_DEFAULT_ENGINE=parallel and get current web results through the ordinary agent search tool without creating a Parallel API key.
2. As an existing operator, I can keep magic or another explicitly configured engine. Adding this option won't change the default or existing choices.

Alternatives Considered

Keep the current Magic search default or configure one of the other supported engines. A separate MCP server entry doesn't make Parallel the provider for the built-in web_search call. Related context: issue #39 raised the Bing-key availability problem and received a suggestion to configure Tavily. This request is for a distinct keyless MCP option, not a default switch.

Implementation Suggestions

The existing MCPClient supports Streamable HTTP. A driver in backend/super-magic/app/tools/web_search_utils/drivers/ could call the MCP web_search tool, map its source URLs, titles, and summaries into the current search result type, and register it under the existing engine selector. Please test provider selection, result mapping, and keyless execution through the existing WebSearch caller.

Priority

Medium - Would be helpful

User Type

Developer (integrating/extending Magic)

Additional Context

The Parallel Search MCP docs say anonymous searches use Fast mode and need no API key. Artificial Analysis' Search API dataset page was updated Sep 22, 2026. Its parallel-fast provider result was last measured Aug 19, 2026: DeepSearchQA average F1 was 0.8024847387112691, and Time per Task was 19.206361937532606 seconds. Those are provider-level results, not an end-to-end Magic benchmark. See the DeepSearchQA chart and the Time per Task chart.

At the baseline, the Super Magic config defaults to magic; the provider selector has no Parallel driver. The current contribution process asks for an issue before a PR.

Btw - I'm a devrel engineer at Parallel.

Checklist

• I have searched for existing feature requests before creating this one
• I have provided a clear description of the proposed feature
• I understand this is a request and may not be implemented immediately