Repository navigation
cloud browser infrastructure for ai agents
Search for "browser infrastructure for AI agents" and every result is a vendor describing itself. That is not a criticism of the vendors; it is just what happens in a young category where the only people writing about the layer are the people selling it. This page is the missing neutral explainer: what the layer is, who is actually in it, what they all have in common, and the question the vendor pages skip, which is whether your setup needs the layer at all.
Disclosure before anything else: this wiki belongs to invisible_playwright_mcp, an open-source agent that ships its own browser and runs locally, which is one of the two answers to that last question. Our interest runs against the category, so read the closing section knowing that, and read the vendor descriptions knowing that every fact in them was checked against that vendor's own site, repository or docs on 2026-09-03.
An AI browser agent has two halves. The agent loop - a model reading a page and deciding the next action - is code you run or a product you use. The browser that loop drives is a heavy, stateful, crash-prone OS process, and "browser infrastructure" is the business of running that process for you, somewhere else.
Concretely, a browser infrastructure vendor gives you an API that starts a browser in their cloud and hands back a connection URL. Your code, unchanged, connects to it over the same protocols it would use locally - Playwright, Puppeteer, or raw CDP - and everything else is their problem: spawning and recycling processes, keeping sessions alive between calls, session recordings and live debug views, egress routing, and scaling from one browser to hundreds. You pay per browser-hour, per credit, or per seat.
That is the whole idea. It is the same move as any managed service: the database did not change when you stopped hosting it, and the browser does not either.
Three facts about browsers make the business real. They are expensive to run at density: a real page can hold a Chromium process at hundreds of megabytes, so a fleet of fifty is a machine-sizing problem, not a loop. They are terrible citizens in CI: no display, missing system dependencies, fonts, and sandbox flags that differ per distro. And agent workloads are bursty: you want zero browsers most of the day and two hundred for ten minutes.
If your agent is a product feature - a SaaS that browses on behalf of your users - those three facts arrive on day one, and renting the layer is usually the right call. Hold that thought against the closing section, because none of the three facts applies to one person running one agent on one machine.
Facts below were read from each vendor's own site, pricing page or repository on 2026-09-03; prices and star counts drift.
Browserbase is the category's mindshare leader: a SaaS with a free tier (3 concurrent browsers, sessions capped at 15 minutes), then $20/mo, $99/mo and custom tiers, its homepage claiming "10,000+ companies". Two things distinguish it beyond scale: it open-sourced Stagehand (MIT, ~24.1k stars), the SDK layer many agents use whether or not they buy Browserbase, and it gives away Director, which turns plain-English instructions into exportable Stagehand code. The open SDK is a funnel to the paid cloud, and it is a good funnel because the SDK is good.
Steel is the credible
open-source answer: Apache-2.0, ~7.6k stars, a browser API you self-host with
a single docker run against their published image, driving Chrome over CDP
with Puppeteer, Playwright and Selenium support. If "open source" in this
category matters to you, this is the project to evaluate first.
Browserless is the veteran: its repository dates to November 2017, years before agents were the customer, and it has ~13.7k stars. Know the license before you build on it: the SPDX line reads "SSPL-1.0 OR Browserless Commercial License". The SSPL is a source-available license the OSI has not approved as open source, so "self-hostable" and "open source" are not the same word here.
Hyperbrowser is a hosted SaaS selling "fast, reliable cloud browsers and sandboxes for AI automation", reachable over Playwright, Puppeteer and CDP, with scraping and crawl APIs alongside the raw sessions.
Anchor Browser is a SaaS whose pitch is the browser build itself: its own Chromium fork it calls "humanized chromium", sold in credit tiers from free (5 credits) through $50, $500 and $2,000 per month, with $1.00 per credit overage. The fork's properties are Anchor's own claims; we did not test them.
Kernel runs isolated Chromium sessions you reach over Playwright, CDP or WebDriver BiDi, and is notable for meeting agents where they now live: it also exposes a hosted MCP server, so a tool-calling assistant can drive its browsers without you writing glue.
Airtop is a SaaS one level up the stack: less "here is a browser" and more an agent-builder, where you describe a workflow in plain language and it compiles it into a deterministic agent that runs on a schedule on their cloud browsers.
Cloudflare is the big-cloud entrant. Its product, recently renamed Browser Run (formerly Browser Rendering), runs headless Chrome on Cloudflare's network, driven over Puppeteer, Playwright, CDP or Stagehand, with REST endpoints for one-shot screenshots, PDFs and markdown. When a company that size treats the layer as a platform primitive, the category has stopped being a startup experiment.
Lightpanda is the odd one out and the most interesting engineering bet on the list: not hosted Chromium but a new headless browser written from scratch in Zig, AGPL-3.0, ~34.4k stars, claiming roughly 9x faster execution and 16x less memory than Chrome on its own benchmark. The honesty is in its README too: it is beta, web API coverage is partial, and some sites will error or crash. A browser for machines, not yet a browser for the whole web.
BrowserStation is the newest self-host option: MIT, tiny today (~166 stars), Kubernetes-native via Ray, each worker pod running a Chrome sidecar that exposes CDP, and explicitly described by its authors as an open-source alternative to Browserbase. Early, but the architecture is the one an infra team would ask for.
One more entrant matters because of where it came from: browser-use, the most-adopted open-source agent framework, now sells "Browser Use Agents" and "Browser Infrastructure" as its two commercial products. The agent side of the market arriving at infra confirms the direction of the money: the layer is where browser-agent companies monetize.
Check the engine column across the list and one fact repeats. Every vendor that names its browser runs Chrome, Chromium or a Chromium fork: Steel and BrowserStation say Chrome, Kernel says Chromium, Anchor is a Chromium fork, Cloudflare says headless Chrome. The vendors that do not name an engine still speak CDP, which is Chromium's protocol. The only exception is Lightpanda, which escaped Chromium by writing a new browser and is paying the price in web compatibility.
Two consequences follow. The convenient one: your automation code is portable, because the whole category standardized on one engine and its protocols, and switching vendors is mostly a connection-string change. The inconvenient one: a cloud browser is the same browser. Moving a stock Chromium automation build from your laptop to a vendor's cloud changes where it runs, not what a page inspecting it sees, and it typically swaps your residential IP for datacenter egress, which is why most of these vendors also sell proxy routing as an add-on. The layer solves operations. It does not solve the questions covered in why does my AI agent get blocked, and the serious vendors do not claim it does.
You need cloud browser infrastructure when the three facts from the top apply: fleets of concurrent browsers, agents running in CI or on servers with no display, bursty scale, sessions that must outlive any one machine, or a product whose users bring the workload. If that is you, the split that matters is managed versus self-host: Browserbase, Hyperbrowser, Anchor, Kernel, Airtop and Cloudflare on one side; Steel, BrowserStation and (with the license caveat) Browserless on the other.
You do not need it when the agent is yours, personal, and local. One person running one agent for their own tasks has no fleet, no CI, and no burst; the laptop that is already on is the infrastructure. This is invisible_playwright_mcp's route, and the disclosure from the top applies in full: invisible_playwright_mcp ships its own browser, a Firefox patched at the C++ level rather than a hosted Chromium, runs it on your machine, and there is no per-hour or per-credit bill because there is no vendor in the loop. Model tokens are the only metered cost. The trade is real in both directions: you get no fleet, no managed scaling and no session cloud, and it is Windows and Linux only. The bill that replaces the per-hour one is disk and memory on your own machine, and it is smaller than the figures this category usually quotes: what a self-hosted AI agent actually costs to run measures it at 613 MB installed and about 900 MB resident, with the model bill kept separate. For what that local route looks like against the rest of the agent field, see open-source AI browser agents.
What is browser infrastructure for AI agents? A managed place for your agent's browser process to run: a vendor API that starts browsers in their cloud and hands your code a Playwright/CDP connection, plus the scaling, session persistence and debugging around it.
What is the best headless browser for AI agents? The engine answer is almost settled: the category runs Chromium, so the real choice is who operates it (the vendors above) or whether you run it yourself. The one non-Chromium bet, Lightpanda, is fast but openly partial on web compatibility today.
Is there an open-source Browserbase? Steel (Apache-2.0) is the mature answer; BrowserStation (MIT, Kubernetes-based) is the new one and describes itself in exactly those words. Browserless self-hosts but is SSPL-licensed, which is source-available, not OSI open source.
Do I need cloud browsers to run an AI agent? No. Every major open-source agent runs on your own machine, and a local agent that ships its own browser (invisible_playwright_mcp is one, disclosure above) has no infra bill at all. The layer earns its cost at fleet scale, in CI, or inside products.
How much does it cost? Entry tiers are cheap ($0 to $99/mo across the SaaS vendors above); the real spend is metered browser-hours, credits and proxy gigabytes at scale, which is exactly when self-hosting starts to look attractive.
See also: Browserbase alternatives for the vendor-by-vendor version of this decision, browser-use alternatives for the agent layer above this one, and AI browser agents vs traditional scraping for whether you need a browser in the loop at all.
All retrieved 2026-09-03. Star counts, prices and tier limits were read from each project's own page on that date and will drift.
- Browserbase, its pricing page, its docs introduction and Director.
- browserbase/stagehand.
- steel-dev/steel-browser.
- browserless/browserless; creation date read from the repository's GitHub API record.
- Hyperbrowser docs.
- Anchor Browser.
- Kernel.
- Airtop.
- Cloudflare Browser Run.
- lightpanda-io/browser.
- ReinforceNow/browserstation.
- browser-use.com, for the two commercial products.
Written while maintaining invisible_playwright_mcp, a local agent that exists so its users do not need this layer. That bias is stated because it is real; the vendor facts above are theirs, checked against their own pages.
- OpenAI Operator alternatives
- Open-source Operator-style agents
- Is OpenAI Operator still available?
- OpenAI Operator vs Claude computer use
- browser-use alternatives
- Choosing an AI browser agent
- Open-source AI browser agents
- Open-source computer-use agents
- What is an AI web agent?
- AI browser agents vs traditional scraping
- Cloud browser infrastructure for AI agents, explained
- Browserbase alternatives
- Firecrawl vs an AI browser agent
- Skyvern alternatives
- Stagehand vs browser-use
- Project Mariner is gone: what replaced it
- Manus alternatives
- Gemini computer use vs Claude computer use
- A stealth browser MCP, reviewed honestly by its own wiki
- AI browser vs AI browser agent: which one do you want?
- AI browser agent vs RPA: which one fits the job
- AI browser agent vs n8n, Zapier and Make
- Vercel agent-browser alternatives, compared honestly
- What is an agentic browser? Definition and the two kinds
- Open-source agentic browsers: the three layers, compared
- Choosing an MCP server for browser automation: four axes
- Stealth MCP servers compared: Camoufox, nodriver, Patchright
- Playwright MCP alternatives, and the three you don't need
- Autonomous browser agents: the four rungs of autonomy
- What is actually free in the AI browser agent stack
- browser-use on GitHub: what the repo actually gives you
- Playwright MCP vs Chrome DevTools MCP: different jobs
- How to choose among MCP servers: a map by category
- Which MCP servers are worth adding to Claude Code
- MCP on GitHub: finding servers and judging them fast
- MCP vs an API: the decision, and what the wrapper costs
- MCP alternatives: when the protocol is the wrong shape
- Why does my AI agent get blocked?
- The timing signal AI agents give off
- Agent retry loops trip rate limits, not fingerprints
- Claude computer use detected as a bot
- browser-use getting blocked: what you can and cannot change
- Playwright MCP session blocked: four causes, four fixes
- Playwright MCP and captchas: what actually gets you past
- Cloudflare and a browser MCP server: what is being read
- Can an AI agent solve a captcha? The honest answer
- Getting an AI agent to fill out forms
- The best model for an MCP browser agent, and what it really costs
- Browser problem or model problem?
- How to use a Playwright MCP server with Claude Code
- Extracting data to a CSV with an AI agent
- Monitoring a page for changes with an AI agent
- How to let Claude Desktop control a browser
- How to add a browser to Cursor as an MCP server
- Using an AI agent to hunt for apartments
- Getting website data into Google Sheets with an AI agent
- Using an AI agent to download invoices from portals
- AI agents for web research
- Using an AI agent to test your own website
- How to add a browser to Cline as an MCP server
- Posting to social media with an AI agent
- Posting to Facebook with an AI agent
- Posting to Instagram with an AI agent
- Posting to X with an AI agent
- Automating LinkedIn posts: read this first
- Appointment bots: what they are and what an agent can legitimately do
- Track prices across sites with an AI agent
- Build a lead list with an AI browser agent
- Run an AI browser agent on a schedule
- AI browser agent with a local LLM: what changes
- Should you log your AI agent into your accounts?
- How to write a task an AI browser agent can follow
- Move data between two web apps with an AI agent
- The MCP server
- How the tools are shaped, and why
- Playwright MCP vs the Playwright CLI: which fits when
- Playwright MCP: browser is already in use, and the fix
- Playwright MCP best practices: four decisions that matter
- Playwright MCP with a proxy, and the three leaks it leaves
- A browser MCP server in GitHub Copilot: setup and limits
- Using a browser MCP server for web scraping: the pattern
- Which LLM for browser automation: the four properties
- How to build a browser agent, and what to take instead
- Getting an AI agent to log into a website: three routes
- MCP tools, resources and prompts: who controls each
- How many MCP tools is too many? The context arithmetic
- How to build an MCP server: the decisions, not the scaffold
- Local or remote MCP server: what changes, and what does not
- Writing an MCP client in Python: the thirty-line version
- Self-hosted AI agent: what one actually costs to run
- How long an AI browser agent takes per step, measured
- Text, HTML, snapshot or screenshot: what the agent should read
- Giving an AI browser agent a stopping condition
- Keeping an AI browser agent out of destructive actions
- Why did the AI agent click the wrong thing
- When the page changes under the AI agent
- Running one AI agent task across a list of sites
- Seeing a page as it appears in another country
- Getting data out of a dashboard with no export button
- Two browsers in one session: main and support
- Finding the dead links on a site with an AI agent
- Filling a CRM record from a company's website
- One form submission per spreadsheet row, with an AI agent
- Dated screenshots of a page as evidence
- Checking order and delivery status with an AI agent
- Reading a PDF that opens inside the browser
- Summarising a long page or thread with an AI agent
- Collecting every image on a page with its caption
- Collecting event and course listings with an AI agent
- Cancelling a subscription with an AI agent
- What an AI agent can and cannot do inside an iframe
- Shadow DOM and an AI agent: you can click it, you cannot read it
- How to add a browser to Codex as an MCP server
- What a page snapshot costs, per control
- How to let Gemini CLI use a browser
- Native selects and the ones that only look like selects
- Clicking by selector or by coordinates
- How long the agent waits before it gives up
- What a second browser costs
- Uploading a file with an AI agent, and why this one cannot
- Watching the agent work, and when it is worth it
- When not to use an AI browser agent
- Agent or script: deciding once instead of every time
- Using the keyboard instead of the mouse
- Secrets in an agent task: where they end up
- What an agent run should log
- Deduplicating what an AI agent collects
- Normalising values across sites
- Validating an AI agent's output
- Reading a table with an AI agent
- Driving a site's own search and filters
- The task works headed and fails headless
- How to let the AdaL CLI use a browser