feat: migrate MCP server search to support Tavily alongside DuckDuckGo

[tavily-integrations]

Summary

Adds Tavily as a configurable search backend for the MCP server's /search and /research endpoints, keeping DuckDuckGo as the default provider.

What changed

• mcp/tools/web.py: Added search_tavily() function using the Tavily Python SDK, with result normalization to {title, href, body} format. Added search_web() dispatcher that routes to Tavily or DuckDuckGo based on the SEARCH_PROVIDER environment variable.
• mcp/server.py: Replaced direct webtools.search_ddg() calls in /search and /research endpoints with webtools.search_web() dispatcher.
• requirements.txt: Added tavily-python>=0.3.0 dependency.
• .env.example: Documented SEARCH_PROVIDER and TAVILY_API_KEY environment variables.

Dependency changes

• Added tavily-python>=0.3.0 to requirements.txt

Environment variable changes

• SEARCH_PROVIDER (optional, defaults to duckduckgo, set to tavily to use Tavily)
• TAVILY_API_KEY (required when SEARCH_PROVIDER=tavily)

Notes for reviewers

• This is an additive change — DuckDuckGo remains the default and is fully preserved.
• The existing search_ddg function is unchanged, so existing tests that mock it directly continue to work.
• Result shape is normalized to {title, href, body} for Tavily results, matching the DuckDuckGo output format.

Automated Review

• Passed after 2 attempt(s)
• Final review: This is attempt 2 addressing two prior review issues. Both fixes are correctly implemented: (1) AsyncTavilyClient is now used with await client.search(), eliminating the blocking-call-in-async-function problem; (2) client instantiation moved inside the try/except block and os.environ.get("TAVILY_API_KEY", "") used for graceful degradation when the key is absent. All relevant files are updated (mcp/tools/web.py, mcp/server.py, requirements.txt, .env.example). Existing tests remain valid — the DDGS mock path still works because SEARCH_PROVIDER defaults to duckduckgo. No regressions detected. One minor doc staleness issue found.

[gemini-code-assist]

Code Review

This pull request introduces Tavily as an alternative search provider alongside DuckDuckGo. It updates the MCP server to use a new dispatcher function, search_web, and adds the necessary configuration and implementation for Tavily searches. Feedback was provided to improve the error handling in the Tavily search implementation by explicitly checking for the API key and ensuring the response processing is safely contained within a try-except block to avoid masking configuration or malformed response issues.

[gemini-code-assist]

The current implementation uses a broad except Exception block that silently returns an empty list, which masks configuration issues (like a missing TAVILY_API_KEY) or API errors. Additionally, processing the response outside the try block could lead to unhandled exceptions if the response is malformed.

Consider checking for the API key explicitly and wrapping the entire search and normalization logic in the try block, ideally with logging for better observability.

Suggested change

	try:
	client = AsyncTavilyClient(api_key=os.environ.get("TAVILY_API_KEY", ""))
	response = await client.search(query=q, max_results=max_results)
	except Exception:
	return []
	results = []
	for r in response.get("results", []):
	results.append({
	"title": r.get("title", ""),
	"href": r.get("url", ""),
	"body": r.get("content", ""),
	})
	return results
	api_key = os.environ.get("TAVILY_API_KEY")
	if not api_key:
	return []
	try:
	client = AsyncTavilyClient(api_key=api_key)
	response = await client.search(query=q, max_results=max_results)
	return [
	{
	"title": r.get("title", ""),
	"href": r.get("url", ""),
	"body": r.get("content", ""),
	}
	for r in response.get("results", [])
	]
	except Exception:
	# Consider adding logging here to capture API or network errors
	return []