How to give Claude Code a browser and which free search is closed?

Briefly

• A real browser opens any page for the agent, but consumes a lot of tokens. Microsoft itself advises agents to use playwright-cli instead of Playwright MCP, because CLI calls are more economical. In the official Anthropic marketplace, there are only browser-based search plugins without keys.
• Google scrapers without a key exist, but everyone is at war with a captcha. Google terms prohibit automatic access, so searching from the main logged-in Chrome profile jeopardizes your Google account.
• Crawl4AI and Firecrawl extract the contents of the pages on their server, but they don't search for them themselves: they are paired with a search MCP server.
• Since 06/18/2026 Gemini CLI does not serve free users and subscribers of Google AI Pro and Ultra. The official way to Google — grounding search (Google search in model responses) in Gemini API: without payment — up to 500 requests per day for Gemini 2.5 Flash and Flash-Lite, and at the free level Google uses your content to improve its products.
• For 2025-2026, the Bing Search API, the free Brave plan, the free OAuth in Qwen Code were closed, and Google Custom Search is closed to new customers. Check the instructions for more than six months with the table at the end of the article.

This is the fourth article of five in the series "Free Internet Search for Claude Code". In the series: built-in WebSearch and WebFetch and MCP connection; services with API and free tariff; search without keys and on your server: SearXNG, DuckDuckGo, multi-engine servers; browsers, scrapers, workarounds and services that have stopped working (this article); search from Russia, security, quality, token consumption and a cheat sheet "what to choose". All data was verified on September 17, 2026.

Claude Code is an Anthropic AI agent for development that works in the terminal and calls the tools itself: reads files, runs commands, searches the Internet. The built-in search is not always enough, and then additional tools are connected to the agent. The MCP server is a program that, using the open Model Context Protocol protocol, gives the agent new tools: search, page loading, browser management.

In this article, we analyze workarounds that do not require a paid search API: a real browser running an agent, Google scrapers, extracting pages on its server and transferring the search to another AI. At the end is a table of services that have stopped working for free. If you are building agents for a team or clients, we collect such solutions as part of the implementation of AI agents.

Is it possible to give Claude Code a real browser?

Yes, and this is the most universal way: the browser opens any page, including the one that is collected by scripts. Minus is the consumption of tokens, that is, units of text that are paid for when working with the model: the page description takes up a lot of space in the context of the agent.

The browser-based MCP server launches Chrome or another browser and gives the agent the commands "open address", "click", "enter text", "read page". A headless browser is a browser without a window that is controlled by a program. CDP (Chrome DevTools Protocol) is the protocol through which an external program controls Chrome.

Microsoft Playwright MCP

Without a key. The server reads the page through snapshots of the accessibility tree, not through screenshots. The accessibility tree is a structured list of page elements (buttons, links, fields) that the browser builds for screen readers; such text is clearer to the model than pictures. Chrome, Firefox, WebKit and Edge are supported, the default profile is permanent, that is, cookies and logins are saved between launches.

The --extension flag connects the server to an already running Chrome or Edge with your sessions. There is also --user-data-dir (path to a separate profile) and --cdp-endpoint (connection to the browser via CDP). Need a Node.js 18 or newer (README playwright-mcp).

Figures: 37.2 thousand stars on GitHub, version v0.0.81 from 09/14/2026.

claude mcp add playwright npx @playwright/mcp@latest

Chrome DevTools MCP

Without a key. The server from the Chrome team officially supports only Google Chrome and Chrome for Testing. By default, it collects usage statistics; it is disabled by the --no-usage-statistics flag (chrome-devtools-mcp repository).

Figures: 52.2 thousand stars on GitHub.

npx -y chrome-devtools-mcp@latest

Claude in Chrome

For a fee. This is the official integration of Anthropic, launched by the claude --chrome command. It uses your browser logins and stops at login pages and captchas so that a person can pass them (Claude Code documentation).

Restrictions:

• need a direct paid plan Anthropic (Pro, Max, Team or Enterprise) and login via /login;
• unavailable when working through Amazon Bedrock, Google Vertex and Microsoft Foundry;
• does not work in WSL (Linux subsystem for Windows);
• increases context consumption if enabled by default.

browser-use (plugin)

Without a key. The plugin manages local Chrome via CDP, you need a Python uv package manager to run it. Cloud browsers in it are paid (README browser-use).

claude plugin install browser-use@browser-use

Websermcp and reference Puppeteer server

Abandoned. BrowserMCP (7.1 thousand stars) has not been updated since 04/24/2025. The reference MCP server on Puppeteer from the official set has been transferred to the archive (servers-archived). The instructions that advise them are outdated.

Summary of browser servers

The tool	Access	Browsers	What is important to know
Playwright MCP	without a key	Chrome, Firefox, WebKit, Edge	accessibility tree; --extension connects your browser with sessions
Chrome DevTools MCP	without a key	Chrome, Chrome for Testing	default statistics, disabled --no-usage-statistics
Claude in Chrome	for a fee (plan Anthropic)	Chrome	not in Bedrock, Vertex, Foundry and WSL
browser-use	without a key (cloud browsers are paid)	local Chrome by CDP	need uv
BrowserMCP, Puppeteer MCP	abandoned	—	Puppeteer MCP is archived, and Browsers have not been updated since April 2025

Why the browser is not the best way to search

Microsoft itself admits that agents increasingly prefer CLI in the form of skills, "because CLI invocations are more token-efficient" — command-line calls are more economical in tokens. Therefore, in the README Playwright MCP, she advises microsoft/playwright-cli (README playwright-mcp).

Playwright, Chrome DevTools and browser-use are available in the official Anthropic plug-in marketplace. There are a total of 308 plug—ins in it, and among the search options, free options without keys are browser-only. Search plugins Exa, Tavily, Bright Data, Firecrawl, Nimble, You.com and Zyte require a vendor account (marketplace.json).

Do Google scrapers work without a key?

They work until Google caught them: each such project describes its struggle with captcha. Scraper is a program that takes data from pages, pretending to be an ordinary visitor, instead of receiving it through the official API.

The project	Status	Stars	What he writes about himself
engram-ae/noapi-google-search-mcp	without a key	128	"Google started catching headless Chromium browsers and serving CAPTCHAs"; the author added stealth patches and a neural network captcha solver, but admits: "Will it solve every CAPTCHA? No» (README)
web-agent-master/google-search and g-search-mcp fork	abandoned since April and June 2025	620 (google-search)	popular, but not updated
HarimxChoi/google-surf-mcp	active	290	sponsored by SearchAPI and offers it as a primary provider (README)
Sophomoresty/gemini-search-mcp	without a key, without a license	—	automates the Google AI Mode tab via Chrome and promises "Free, unlimited", but documents the captcha /sorry/, which is solved manually (README)
mixelpixx/Nimrod	for a fee	—	turned into a paid service with trial 50 credits (README)

Stealth patches are edits that hide the signs of an automated browser from the site. A project without a license cannot formally be freely used and modified: the rights remain with the author. This is a separate risk for the company's work environment.

How will Crawl4AI and Firecrawl help on your server?

They extract the content of the pages in a form convenient for the model, but they do not search themselves. They are paired with a search MCP server: the search finds addresses, and the crawler picks up the text. A crawler is a program that crawls pages and saves their contents. Self-hosted means that the service runs on your server, and not in the vendor's cloud.

Crawl4AI

Without a key. It is launched by a single Docker container and immediately gives MCP (Docker README Crawl4AI). Tools: md (page in Markdown), html, screenshot, pdf, execute_js (execute JavaScript on the page), crawl and ask. There is no web search tool among them (README Crawl4AI).

Connection is via SSE (Server-Sent Events) — outdated MCP transport: in the protocol specification version 2025-03-26, the HTTP+SSE transport has been replaced with Streamable HTTP (MCP specification: Transports), and Claude Code marks --transport sse as obsolete.

Numbers: 83.8 thousand stars on GitHub, version v0.9.3, port 11235.

docker run -d -p 11235:11235 --name crawl4ai --shm-size=1g unclecode/crawl4ai:latest
claude mcp add --transport sse c4ai-sse http://localhost:11235/mcp/sse

The docker run command publishes port 11235 on all network interfaces of the machine. On a server accessible from the network, the port should be linked to 127.0.0.1 or closed with a firewall, otherwise anyone can use the execute_js tool.

Firecrawl self-hosted

Without a key, AGPL-3.0 license, port 3002. AGPL-3.0 is a copyleft license: if you have changed the code and give access to the service over the network, you need to open the modified code. The stack is heavy: API, workers, Playwright, Redis, RabbitMQ, PostgreSQL and FoundationDB (SELF_HOST.md ).

What is important to know:

• by default, the server works without authentication, that is, anyone who reaches the port can use it;
• screenshots, page actions, and Agent and Browser functions are not available. They are only available in the cloud version (Firecrawl documentation);
• The Firecrawl MCP server connects to its installation via the FIRECRAWL_API_URL variable without a key (README firecrawl-mcp-server);
• It was not possible to find out how the self-hosted version performs /search without a foreign key: only SearchAPI variables (.env.example) are visible in .env.example.

It is more convenient to keep both services on a separate server or virtual machine, rather than on the developer's laptop. For such tasks, we prepare the server infrastructure and close access to service ports.

Is it possible to give the search to the free Gemini CLI?

No, this path has been closed for free accounts since 06/18/2026. The most famous trick of 2025 looked like this: Claude Code calls gemini -p, and he searches through his google_web_search tool. This no longer works for free accounts.

Google announced on 05/19/2026 that from 06/18/2026 Gemini CLI ceases to serve free users, as well as subscribers of Google AI Pro and Ultra. Access was retained by owners of Gemini Code Assist Standard and Enterprise licenses and users of paid API keys (Google Developers Blog; discussion #27274).

The official quota page confirms the replacement with Antigravity CLI, but below it still shows "1000 requests per day" and "250 per day for a free key" (geminicli.com ). These figures are outdated and contradict the announcement.

Back in March 2026, Google threatened with sanctions for "using Gemini CLI OAuth with third-party software" (discussion #22970). OAuth is an account login protocol in which the program gets access on your behalf. The ban also applies to wrappers that transfer a personal OAuth session to another program.

Antigravity CLI (agy)

Free tariff with weekly quota updates. Antigravity CLI (agy team) is the successor to Gemini CLI. For users without AI Pro and Ultra, it gives "Meaningful quota, refreshed weekly" — "a tangible quota that is updated weekly", without specific figures (Antigravity Plans). The remainder is shown by the /usage command (Antigravity documentation).

Unofficial sources say about 20 requests per day and weekly blockages (not confirmed by the original source; antigravity.im ). The request of users to return the Gemini CLI quota (issue #79) indirectly confirms that it is now less.

MCP wrappers over Antigravity

• jamubc/gemini-mcp-tool (2283 stars) uses Antigravity by default, and for the paid key there is a variable GEMINI_MCP_BACKEND=gemini (gemini-mcp-tool).
• delegate-agy (16 stars) adds an agy-delegate-search subagent to search with quotes (delegate-agy). Subagent — a separate instance of the agent to which the main one transfers the subtask.

claude mcp add gemini-cli -- npx -y gemini-mcp-tool

That agy has a grounded search tool at all is claimed only by the README of these projects (according to the project's README). Plug—ins on the old Gemini CLI - d-oit/gemini-search-plugin, mcp-gemini-web-search, choplin/mcp-gemini—cli - for free users are actually broken.

Gemini API: grounding via Google Search

Free tariff with a daily limit, but only for some models. Grounding is a mode in which the model searches Google before responding and relies on the pages found. According to the Gemini API pricing page (verified 09/17/2026), the conditions depend on the access level:

Models	Free Tier	Paid Tier
Gemini 2.5 Flash and Flash-Lite	up to 500 grounded requests per day, the limit is shared between the two models	1500 requests per day for free, then $35 for 1000
Gemini 2.5 Pro	not available	1500 requests per day for free, then $35 for 1000
Gemini 3.x (most models)	not available	5000 requests per month for free on all 3.x models, then $14 for 1000

The figures of 1500 per day and 5000 per month are often retold as a free tariff, but this is a free volume within the paid level: it requires connected billing. Without payment, grounding is available only on Gemini 2.5 Flash and Flash-Lite (as well as on Live models for voice agents).

At the free level, Google uses your content to improve its products, but not at the paid level. Therefore, it is better not to send requests with customer data or internal information to the free level.

MCP wrappers for this mode are still small: claudiusbotticus/gemini-research-mcp (17 stars, without license) and shariqriazz/google-ai-search-mcp (6 stars, MIT license, available in GitHub MCP Registry) (README gemini-research-mcp; GitHub MCP Registry).

Other "free" delegates are less reliable

Service	What is in fact	Source
Search Z.ai (GLM)	only for GLM Coding Plan subscribers	docs.z.ai
GrokSearch (1855 stars)	requires GROK_API_KEY	README
ContextX Public Grok Endpoint	it works without authorization, but it is an anonymous proxy with an undisclosed source and without a license	README
Kagi-Session2API	he himself warns about the possible ban of the Kagi account	README
Codex CLI (OpenAI)	sources differ: the official pricing page refers the CLI to Plus plans and higher, third-party articles write about a small access to Free. It was not possible to officially confirm whether he has a search	learn.chatgpt.com ; eesel.ai (secondary source)

An anonymous proxy is dangerous not only because it can disappear. Everything that the agent sends there is seen by an unknown owner, and everything that comes back may contain hidden instructions for the model. In a working environment, such tools should not be connected; we understand how to assess these risks in the work on information security.

Which free search services closed in 2025-2026?

Bing Search API, the free Brave plan, the free Gemini CLI and the free OAuth in Qwen Code have been closed for a year and a half, and Google Custom Search no longer accepts new customers. Therefore, any instruction older than six months should be checked with this table.

Service	Status as of 09/17/2026	Replacement or bypass	Source
Bing Search API, including the free F1 tariff	closed: disabled 08/11/2025	paid Grounding with Bing in Azure AI Agents	Microsoft Learn
Google Custom Search JSON API (100 requests per day for free)	closed for new clients; open for current clients until 01.01.2027	—	Google for Developers
Free Brave Search API plan (2000, then 5000 requests per month)	closed in February 2026 (according to a secondary source)	card binding and $5 credits	implicator.ai
Gemini CLI for free users, Google AI Pro and Ultra	not serviced since 06/18/2026	Antigravity CLI	Google Developers Blog
Qwen Code, free login via OAuth	the quota was cut from 1000 to 100 requests per day, then closed on 04/15/2026	—	Decrypt; issue #3316
$5 per month of Perplexity API Credits for Pro Subscribers	The status is unclear: according to third-party data, removed in early 2026; not officially confirmed	—	goosed.ie (secondary source)
Reference Puppeteer MCP and BrowserMCP	abandoned: the first is archived, the second has not been updated since 2025	Playwright MCP, Chrome DevTools MCP	servers-archived; BrowserMCP
DuckDuckGo via regular httpx	broken: since June 2026 gives an empty response with the code 202	installation with the [browser] add-on	issue #46

httpx is a Python library for HTTP requests; many MCP servers for DuckDuckGo work through it. Code 202 means "request accepted", but in this case the response body is empty and the server does not receive results. For more information about DuckDuckGo and keyless search, see the previous article in the series.

The table shows that the free tariffs are closed faster than the instructions for them are updated. Cloud free tariffs that are still working are discussed in the second article of the series. For tasks that should work for years, it is more reliable to rely on your own search engine installation, rather than on someone else's generosity.

What is the workaround to choose for a work task?

For regular search — a search MCP server, and a browser and a crawler - only as an add—on. Below is the order that follows from the facts of this article.

1. Need a search. Connect a search MCP server: cloud-based with a free tariff or your own SearXNG. Google scrapers and anonymous proxies are not suitable for this: they break on the captcha and violate the terms of the search engine.
2. The text of the found pages is needed. Add Crawl4AI to Docker on your server and close its port from the external network.
3. The page requires login or action. Connect the Playwright MCP with a separate profile via --user-data-dir. If you have a direct paid Anthropic plan, Claude in Chrome is also suitable, which gives the input and captcha to a person.
4. I want a free Google. The official way is grounding in the Gemini API: at a free level, up to 500 requests per day for Gemini 2.5 Flash and Flash-Lite, but only for requests without sensitive data, because Google uses this content to improve its products. With billing enabled, the free volume is larger (1500 per day for 2.5, 5000 per month for 3.x), and the content for training is not used.
5. We found the instructions on the Internet. Check it with the table of closed services and the date of the last update of the repository.

How to organize a search from Russia, protect the agent from prompt injection — harmful instructions hidden in the texts of the pages — and how many tokens each method spends, we will analyze in the next article of the series. If an agent needs to be embedded in your systems, this is a development task: we select tools for the limitations of a particular company.

Sources

1. Microsoft Playwright MCP, README — github.com , verified on 09/17/2026
2. Chrome DevTools MCP, repository — github.com , verified on 09/17/2026
3. Claude Code: Use Claude Code with Chrome — code.claude.com , verified on 09/17/2026
4. browser-use plugin, README — github.com , verified on 09/17/2026
5. Browsers, repository — github.com , verified on 09/17/2026
6. Archive of the reference Puppeteer MCP server — github.com , verified on 09/17/2026
7. The official marketplace of Anthropic plugins, marketplace.json — raw.githubusercontent.com , verified on 09/17/2026
8. noapi-google-search-mcp, README — raw.githubusercontent.com , verified on 09/17/2026
9. google-surf-mcp, README — raw.githubusercontent.com , verified on 09/17/2026
10. gemini-search-mcp, README — raw.githubusercontent.com , verified on 09/17/2026
11. Nimrod, README — raw.githubusercontent.com , verified on 09/17/2026
12. Google Terms of Service — policies.google.com , verified on 09/17/2026
13. Crawl4AI, README — github.com , verified on 09/17/2026
14. Crawl4AI, Docker README — github.com , verified on 09/17/2026
15. Firecrawl, SELF_HOST.md — github.com , verified on 09/17/2026
16. Firecrawl: self-hosting — docs.firecrawl.dev, verified on 09/17/2026
17. firecrawl-mcp-server, README — github.com , verified on 09/17/2026
18. Firecrawl, .env.example — github.com , verified on 09/17/2026
19. An important update: transitioning Gemini CLI to Antigravity CLI — developers.googleblog.com , verified on 09/17/2026
20. Gemini CLI, discussion #27274 — github.com , verified on 09/17/2026
21. Gemini CLI: quota and pricing — geminicli.com , verified on 09/17/2026
22. Gemini CLI, discussion #22970 — github.com , verified on 09/17/2026
23. Antigravity: Plans — antigravity.google, verified on 09/17/2026
24. Antigravity CLI: command /usage — antigravity.google, verified 09/17/2026
25. Antigravity limits (unofficial website) — antigravity.im , verified on 09/17/2026
26. Antigravity CLI, issue #79 — github.com , verified on 09/17/2026
27. gemini-mcp-tool, repository — github.com , verified on 09/17/2026
28. delegate-agy, repository — github.com , verified on 09/17/2026
29. Gemini API pricing — ai.google.dev, verified on 09/17/2026
30. Model Context Protocol, specification 2025-03-26: Transports — modelcontextprotocol.io , verified on 09/17/2026
31. gemini-research-mcp, README — raw.githubusercontent.com , verified on 09/17/2026
32. GitHub MCP Registry, search by word search — github.com , verified on 09/17/2026
33. Z.ai: Search MCP Server — docs.z.ai , verified on 09/17/2026
34. GrokSearch, README — raw.githubusercontent.com , verified on 09/17/2026
35. ContextX, README — raw.githubusercontent.com , verified on 09/17/2026
36. Kagi-Session2API-MCP, README — raw.githubusercontent.com , verified on 09/17/2026
37. ChatGPT: pricing — learn.chatgpt.com , verified on 09/17/2026
38. OpenAI Codex free access explained — eesel.ai , verified on 09/17/2026
39. Bing Search API retirement — learn.microsoft.com , verified on 09/17/2026
40. Custom Search JSON API: overview — developers.google.com , verified on 09/17/2026
41. Brave drops free Search API tier — implicator.ai , verified on 09/17/2026
42. Alibaba shuts down free tier of Qwen Code — decrypt.co , verified on 09/17/2026
43. Qwen Code, issue #3316 — github.com , verified on 09/17/2026
44. Perplexity Pro quietly removes free API credits — goosed.ie , verified on 09/17/2026
45. duckduckgo-mcp-server, issue #46 — github.com , verified on 09/17/2026