All insights

Coding agents

Claude Code Web MCP: Tavily Search, Chrome DevTools, and WebMCP

Compare Claude Code options for web search and browser control, set up and verify Chrome DevTools MCP, and understand its experimental WebMCP support and production API distinctions.

5 min readTalk to our API team
Illustration for Claude code web mcp

“Claude Code web MCP” can refer to different tools. For web search, Tavily documents a remote MCP server with a tavily_search tool. For browser control and debugging, Chrome DevTools MCP connects Claude Code to a live Chrome browser. WebMCP means structured tools provided by a webpage; Chrome DevTools MCP has experimental support for discovering and invoking those tools, but that is separate from native WebMCP support in Claude in Chrome. (Tavily MCP docs) (Chrome DevTools setup) (Chrome WebMCP debugging)

Checked November 10, 2026. Choose according to the task: search the web, control a browser, or interact with tools a webpage exposes. The setup and availability details below apply to the specific product surfaces described—not every Claude product or MCP server.

Which “web MCP” do you need?

What you want to do Option What to know
Search the web for current sources Tavily MCP Tavily documents a tavily_search tool and a remote HTTP setup for Claude Code. Its instructions say to start a new session and complete OAuth sign-in. Tavily also says Search and Extract are available through keyless access. (Tavily MCP docs)
Navigate pages, interact with controls, or inspect a live browser Chrome DevTools MCP Chrome documents Claude Code as a supported MCP client. The server provides browser automation and DevTools capabilities; it is not a dedicated web-search tool. (Chrome DevTools setup)
Discover or invoke structured tools published by a webpage WebMCP through Chrome DevTools MCP The server documents experimental WebMCP tools. They are not active by default and require the applicable Chrome WebMCP support plus an experimental server option. This does not establish native WebMCP support in the separate Claude in Chrome extension. (WebMCP debugging guide) (Server configuration)

Quick choice: use Tavily for search; use Chrome DevTools MCP for live browser interaction; use its experimental WebMCP tools when testing a site that exposes WebMCP. Check the required browser and server settings before relying on WebMCP for a task.

Set up web search with Tavily

Tavily’s documented Claude Code command is:

claude mcp add tavily --transport http https://mcp.tavily.com/mcp/

Run it in a terminal. Start a new Claude Code session and complete the browser-based OAuth sign-in. Tavily documents tavily_search for web search and says its keyless access supports Search and Extract without an API key. (Tavily MCP docs)

Set up Chrome DevTools MCP in Claude Code

Chrome’s setup guide lists Node.js (latest LTS), npm, and current stable Chrome as prerequisites. In a Claude Code session, add the marketplace and open the plugin installation flow: (Chrome DevTools setup)

  1. Add the marketplace:

    /plugin marketplace add ChromeDevTools/chrome-devtools-mcp

  2. Start installation:

    /plugin install chrome-devtools-mcp@chrome-devtools-plugins

    Review the plugin details and choose an install scope in the plugin panel to complete installation. Claude Code’s plugin documentation says installing a plugin adds its MCP servers to Claude Code. (Claude Code plugin installation)

  3. Run /mcp and check the Chrome DevTools server’s connection status. Claude Code’s MCP quickstart describes checking server status as part of verifying setup. (Claude Code MCP quickstart)

  4. Test an actual, read-only browser call. Ask Claude: “Use Chrome DevTools MCP to navigate to https://example.com, then take a page snapshot and report the page title. Do not click, submit, or change anything.” Count the test as successful only if the session shows a navigate_page call from Chrome DevTools MCP and its tool result, followed by the page snapshot—not just an answer Claude could provide from memory. The server’s tool reference documents navigate_page and take_snapshot. (Chrome DevTools MCP tool reference)

The plugin listing documents browser interactions, screenshots, DOM snapshots, performance analysis, network inspection, console messages, and Lighthouse audits. Chrome warns that the agent can inspect and modify browser data, and that an authenticated session can let it act on your behalf. Use a low-risk browsing context and review actions before allowing changes. (Anthropic plugin listing) (Chrome DevTools setup)

What about WebMCP tools on a webpage?

Chrome DevTools MCP’s current documentation lists list_webmcp_tools and execute_webmcp_tool, but marks WebMCP support as experimental and disabled by default. Chrome’s guide says WebMCP requires Chrome 149 or later with WebMCP enabled—for example, through the applicable origin trial or a local Chrome flag—and the MCP server must be started with --categoryExperimentalWebmcp. Check the current server and browser documentation for exact configuration before trying it. (Chrome DevTools MCP tool reference) (WebMCP debugging guide) (Server configuration)

This is support through the Chrome DevTools MCP server’s experimental tools, not evidence that Claude in Chrome itself natively discovers webpage-registered WebMCP tools. A GitHub issue requesting WebMCP support in the Claude Chrome Extension is marked “closed as not planned”; that issue status does not determine support in other integrations or establish the current state of every Claude product. (GitHub issue #30645)

Claude Code MCP is different from a production model API

A claude mcp command configures tools for a Claude Code client; it is not a model API request. Anthropic documents a separate MCP connector for the Messages API, with remote MCP server definitions and tool configuration. That connector supports remote HTTP servers using Streamable HTTP or SSE, but does not directly connect to local stdio servers. (Claude Code MCP docs) (Messages API MCP connector)

  • Interactive coding session: add an MCP server or plugin to Claude Code.
  • Application backend calling Claude: use the Messages API; use its MCP connector if the application needs a remote MCP server in that workflow.
  • Local stdio server: connect it through a compatible local MCP client; it is not directly connectable through the Messages API MCP connector. (Messages API MCP connector)

Keep client-side MCP configuration separate from production API credentials and requests. For endpoint, model, authentication, request format, and connector requirements, follow the current Messages API documentation rather than copying a Claude Code CLI command into an application. (Claude API overview)

Bottom line

For web search, Tavily documents a search tool and a Claude Code remote-MCP setup. For browser control and debugging, Chrome DevTools MCP provides a documented Claude Code plugin route; verify the server status and a visible tool call. For WebMCP, Chrome DevTools MCP documents experimental tools with additional browser and server requirements—do not confuse that with native support in Claude in Chrome. For a production application, use the Messages API and its remote MCP connector when appropriate, not a Claude Code MCP setup command.

Sources

Tavily MCP

Get started with Chrome DevTools for agents

Chrome DevTools plugin for Claude

Install and manage plugins

Connect to MCP servers

Chrome DevTools MCP tool reference

Debug WebMCP tools with AI agents

Chrome DevTools MCP configuration

WebMCP support in Claude Chrome Extension — issue #30645

Connect Claude Code to tools via MCP

MCP connector

API overview

Contact us