The Best Browser Agent MCP Servers (2026)
By Steven Van ·
Browser Use compares four MCP servers that run a whole browser agent from one goal, covering models, cost, live view and when to pick each.
Browser Use has published a comparison of MCP servers that run a whole browser agent from a single goal, with Browser Use listed first. The post covers its own Cloud and open-source servers alongside Hyperbrowser and agent-browser. Browser Use says it makes Browser Use and Browser Harness, so they come first, and that competitor facts come from their own repos, docs and pricing pages, checked October 18, 2026.
Delegate or drive
Driving servers (Playwright MCP, Chrome DevTools MCP, Browser Harness) expose clicks, typing and page reads, so the calling agent plans every step and reads every page. Delegating servers expose one tool that takes a goal in plain words. A second agent with its own model runs the browser and returns the result.
In Browser Use's matched eval (100 BU Bench V2.1 tasks, Codex with GPT-6 Luna, one run per task), four driving tools used 2.1M to 6.7M median input tokens per task. None of the delegating servers were in that eval, so the post gives no figure for the savings.
Browser Use Cloud MCP
This is a hosted server at https://api.browser-use.com/v3/mcp that runs the Browser Use agent on cloud browsers, with no model key needed. It has seven session tools: run_session, send_task, get_session, get_session_messages, stop_session, list_sessions and list_browser_profiles.
- run_session takes a task of up to 50,000 characters and returns right away. The agent then polls get_session for status, step count, output, cost and a live URL.
- output_schema returns the output as JSON in a given shape. With keep_alive: true, send_task can send follow-ups to the same browser.
- The model is chosen per task from 11 options, including gemini-3-flash, gpt-5.4-mini, claude-sonnet-5 and claude-opus-4.8.
- Sessions use a US residential proxy by default. proxy_country accepts any country code, and null turns the proxy off. profile_id starts a task already logged in with a saved cloud profile.
- New projects can run 10 sessions at once. $100 paid unlocks 50, and spend tiers go up to 1,000.
Pay-as-you-go pricing is $0.02 per browser-hour, billed by the minute, model tokens at provider price plus 20%, and residential proxy at $5/GB ($0.20/GB for direct traffic with the proxy off). The post puts a 10-minute run at about a third of a cent for browser time.
The post lists limits: it runs API V3, which Browser Use's docs say trades accuracy for speed and lower cost than V4, and GPT-6 Luna is not in the MCP model list. There is no mid-task click, no use of local Chrome, and recording is only on the REST API. For accuracy-critical new integrations, the post recommends calling API V4 directly.
Browser Use open-source MCP
The MIT-licensed library runs as a local stdio MCP server with 16 tools. Fifteen are direct browser control. The sixteenth, retry_with_browser_use_agent, runs a full agent on a plain-words task, for up to 100 steps by default, with an allowed_domains list.
It needs Python 3.11 or newer and your own model access: an OpenAI key, an OpenAI-compatible endpoint, or AWS Bedrock. The post says the docs mention an Anthropic key but the server code does not read it. It is free apart from model costs, and has no stealth, managed proxies or CAPTCHA help. Because the browser runs locally, it can reach localhost, VPN and staging sites. Anonymous telemetry is on by default and can be turned off with ANONYMIZED_TELEMETRY=false.
Other servers
Hyperbrowser's MCP server is an MIT Node server with 10 tools. Three are agents: browser_use_agent, openai_computer_use_agent and claude_computer_use_agent. Browser Use says it is the only other server in its round-up with whole-task agents as MCP tools. agent-browser has an opt-in chat tool for short instructions.