Local agent blocks could use web search to find current public datasets, check documentation, and cite sources in notebook analyses. Parallel Search MCP provides web search and page fetching without a Parallel account or API key.

I propose extending the MCP configuration in @deepnote/blocks and the loader in packages/runtime-core/src/agent-handler.ts to support HTTP MCP, then making Parallel available when search is unconfigured. Existing project/block servers and overrides would retain their behavior, with an explicit search disable control and no fallback after a configured provider fails.

I would test discovery, search and fetch through the agent tool loop, result consumption, configuration precedence, cancellation, timeouts, and errors, and update the affected package, product, and skill references.

Would this fit the local agent-block roadmap? I work at Parallel.