Claude Code on Bedrock and Vertex AI in 2026: Web Search Availability and Workarounds

TLDR: Claude Code has no built-in web search on Amazon Bedrock as of September 2026, per Anthropic's feature availability documentation. On Google Cloud's Agent Platform, formerly Vertex AI, the built-in tool works for Claude 4 models and later. Anthropic's own docs say the search backend is not configurable and that the way to search with a different provider is to add an MCP server that exposes a search tool. The You.com Web Search API is available as a remote MCP server, so one claude mcp add command gives Claude Code the same grounded web search on Bedrock, Vertex AI, and the Anthropic API.

Teams deploy Claude Code through a cloud provider for good reasons: AWS or GCP billing, IAM policies, and audit logs in the account they already run. The tradeoff is that some Claude Code features do not follow the model to the new provider, and web search is the one that surprises people most. This guide covers exactly what is available where, why the gap exists, why a search layer you control is the better fix even where the built-in tool exists, and how to wire it in.

Which Claude Code Deployments Include Web Search?

Anthropic publishes a per-provider feature matrix for Claude Code. As of September 2026, the web search row reads as follows (Anthropic feature availability documentation, fetched 2026-09-04):

So the claim that Claude Code has no native web search on Bedrock and Vertex is half right. Amazon Bedrock is a clean gap. On Vertex AI, now branded Google Cloud's Agent Platform, the tool exists with a model-version caveat: if your deployment pins Claude 3.x models, you will not see web search. Note the Claude Platform on AWS row, too. It bills through AWS Marketplace but calls the Anthropic API directly, which is why it keeps web search while Bedrock does not.

Why Is Web Search Missing on Amazon Bedrock?

Anthropic's documentation groups web search with CLI features that "depend on a server-side capability that not every provider exposes" (feature availability, 2026-09-04). The built-in WebSearch tool is not something the Claude Code CLI runs locally. It calls Anthropic's web search backend on the inference side, and when your requests route through Amazon Bedrock, that capability is not part of the request path. The Bedrock setup guide states it plainly: "The WebSearch tool is not available on Amazon Bedrock."

The practical symptom: everything else works. File editing, shell commands, codebase search, subagents, and any MCP servers you configure yourself. But the model's tool list never includes a built-in web search entry, and questions that need current information get answered from training data with no citation.

Anthropic's suggested alternative on Bedrock is "the WebFetch tool with a specific URL." That helps when you already know the page. It does not help when the question is which page, and the tools reference adds that WebFetch is lossy by design: it runs an extraction prompt through a small model and hands Claude that model's answer rather than the raw page. Search is discovery. Fetch is reading. A grounded coding session needs both.

Does Claude Code on Vertex AI Have Web Search?

Yes, for Claude 4 and later models, per the September 2026 feature matrix. Two setup details from Anthropic's Agent Platform guide still matter. First, pin your model versions with ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_OPUS_MODEL, and the related variables. Without pinning, aliases such as sonnet resolve to Claude Code's built-in default for the provider, which can lag the newest release and may not be enabled in your project. Second, if web search silently never fires, check that the pinned model is actually a Claude 4 or later model, because the capability tracks the model generation.

Even where the built-in tool works, it has fixed limits: it returns result titles and URLs and not page content, and the backend cannot be changed. That is the reason to read the next section even if you are on Vertex AI and not on Bedrock.

Why Ground Claude With a Search Layer You Control?

Grounding means giving the model current, cited evidence at answer time instead of letting it reason from training data. Coding sessions are the worst case for ungrounded answers: library versions, deprecated flags, renamed APIs, and breaking changes move faster than any training cutoff, and a model that answers from memory does so with full confidence. On Bedrock there is no built-in grounding at all. On Vertex AI there is a built-in tool you cannot configure. In both cases a search layer you control is the stronger position, for five reasons.

One configuration across every provider. Anthropic lists MCP servers under features available on every provider. The MCP client runs on the developer's machine and connects to the search server directly, so the same .mcp.json works for a developer on the Anthropic API, a team on Amazon Bedrock, and a team on Agent Platform. Nobody is second class, and Anthropic's enterprise guidance already recommends that one central team check a .mcp.json into the codebase.

The same results regardless of inference provider. Because the search call never touches the model provider, a prompt that works in one deployment returns the same evidence in another. That makes agent behavior reproducible across environments and makes evals comparable, which the built-in tool cannot promise when it exists on one provider and not the next.

Control over what the agent is allowed to read. The You.com Web Search API accepts include_domains, exclude_domains, and boost_domains, each up to 500 domains (Search API reference, 2026-09-04). A platform team can pin research to the documentation sites it trusts, or exclude sites it does not, at the request level. Freshness filters (day, week, month, year, or a date range) keep answers on current releases.

Evidence the agent can cite. Every result carries a URL, title, description, snippets, and a page_age timestamp, and web and news results arrive in one response. When the agent quotes a release note, you can see which page it came from and how old it is. The you-research tool goes further with synthesized, citation-backed answers when a question needs more than a result list.

Cost you can see. The Web Search API is priced at $5.00 per 1,000 calls, full page extraction and the Contents API at $1.00 per 1,000 pages, and new accounts start with $100 in credits with no card required (You.com quickstart, pricing section, 2026-09-04). A free profile of the MCP server needs no signup at all, which makes the next section a zero-commitment test.

How Do You Add Web Search to Claude Code on Bedrock?

MCP, the Model Context Protocol, is Claude Code's extension mechanism for external tools, and it runs client side. You.com runs a remote MCP server at https://api.you.com/mcp that authenticates with OAuth 2.1 or a bearer API key (You.com MCP server documentation, fetched 2026-09-04). The install is one command, and Claude Code opens a browser window for the OAuth sign-in automatically:

claude mcp add --transport http you-com https://api.you.com/mcp

For CI, scripts, or shared dev containers, pass an API key from the You.com platform as a header instead:

claude mcp add --transport http you-com https://api.you.com/mcp \
  --header "Authorization: Bearer $YDC_API_KEY"

To try it with no account and no key, use the free profile, which exposes you-search only and allows 100 queries per day:

claude mcp add --transport http you-com "https://api.you.com/mcp?profile=free"

Each command writes to local scope by default. To share the server with everyone who clones the repository, add --scope project, or commit this .mcp.json at the project root. Team members approve it the first time they run claude in the project:

{
  "mcpServers": {
    "you-com": {
      "type": "http",
      "url": "https://api.you.com/mcp"
    }
  }
}

Verify the connection with claude mcp list, which shows a health status next to each server, then check status from inside a session with /mcp. You should see you-com listed with you-search, you-contents, and you-research among its tools (Anthropic MCP documentation and You.com Claude Code guide, 2026-09-04). The commands are identical on Bedrock, Agent Platform, and the Anthropic API.

What Does the You.com Search Tool Return Inside Claude Code?

Once connected, you-search appears in the model's tool list like any other tool, and Claude calls it when a question needs current information. The Web Search API behind it returns web and news results in one response, each carrying a URL, title, description, snippets, publication date, thumbnail, and favicon (You.com Search API documentation, fetched 2026-09-04). When snippets are not enough, the extraction_mode option returns query-relevant highlights from each page, sized for agents that run several searches per task.

Search returns the source list. When the agent needs the full page, you-contents on the same server fetches URLs and returns Markdown, HTML, or metadata, with a max_age control that bypasses the cache when set to 0 (You.com Contents API documentation, fetched 2026-09-04). The common chain is search for the current guide, then pull the full page for the one result that matters.

What Breaks, and How Do You Detect It?

The failure mode to know: the server is configured, the model never calls any search tool, and the session quietly answers from training data. You asked for this week's library release notes and got a confident recap of two versions ago. Detection is a test prompt that cannot be answered from memory. Ask for something posted this week, then check that the session reports a tool call rather than a bare answer.

The second failure mode is stale configuration. The API key rotated, the header still carries the old one, and every search call fails auth, which looks identical to not being configured at all. If the test prompt produces no tool call, run claude mcp get you-com. A failed server shows an Issue: line with the HTTP status the server returned, so a 401 points at the key before you debug anything else (Anthropic MCP documentation, 2026-09-04).

Beyond Claude Code: Grounding Claude on Bedrock and Vertex AI in Your Own Agents

The same gap appears in agents you build yourself. Claude on Amazon Bedrock and Claude on Agent Platform support tool use, but the search tool has to come from you. The You.com Web Search API is a single HTTP call with a JSON body, so it drops into a tool definition in any framework, and You.com publishes Python and TypeScript SDKs plus the remote MCP server for frameworks that speak MCP directly. The result is one grounding layer for both the interactive coding sessions and the production agents, with the same domain controls and the same cited results. For the general pattern, see the LLM web search API guide.

Where Does This Fit With the Rest of the Tooling?

This guide covers the provider-specific gap. The general install, the built-in WebSearch and WebFetch limits on the Anthropic API, and the prompt injection risk that comes with any tool that reads external content are in the Claude Code web search tool guide. The live web search API guide covers when a dedicated search layer beats built-in model search.

Next action: run the free-profile install, ask Claude Code for one piece of information from this week, and confirm the tool call shows up in the session. Then get an API key from the You.com platform and commit the .mcp.json so the whole team gets grounded search on Bedrock, Vertex AI, or wherever Claude Code runs next. Current rates are on the You.com billing page.