mrkeyoor.com_
Tue 01 Sept 16:13 UTC
BrowsersMCP servermeasured 19 Aug 2026

Browserbase

Browserbase MCP is the Browserbase-built bridge between an AI client and a cloud Chrome session running on their infrastructure. It exposes six tools: start, navigate, act, observe, extract, end. Instead of driving a local browser, your agent rents a remote one, so page loads, captchas and proxies happen on Browserbase hardware rather than your laptop.

BootBOOTS ✓We started it and initialize came back in 21.7s.
Tools6tools/list payload 1,570 bytes
Context tax~393estimated tokens for the whole tools/list, chars/4 estimate
Resources / prompts0 / counts returned by resources/list and prompts/list
Measured19 Aug 2026one run of the harness, clean environment, no credentials · how we test

How we started it

npx -y @browserbasehq/mcp
Verdict

Use it if you already pay Browserbase, or intend to. The tool surface is the smallest useful browser API we measured this batch: six tools, tools/list at 1570 bytes, roughly 393 tokens by the chars/4 estimate (measured 2026-08-19). The cost against it is the boot: 21702 ms from spawn to initialize on a cold npx cache, and it starts happily with dummy credentials, so a misconfigured setup looks alive right up until the first navigate fails.

Use it if

You want browser automation your agent can call without a Chrome binary, X server or headless-shell install on the machine running the client
Your scraping targets block datacenter IPs and you were going to buy proxy infrastructure anyway
You are keeping a tight context budget and cannot spend thousands of tokens on a browser tool schema before the agent does any work
You prefer act/observe/extract semantics over raw click-by-selector, and accept the model calls that hide behind them

Skip it if

You will not send credentials to a third party; every page your agent opens is loaded inside someone else's browser
You want free local automation, in which case a Playwright-based server costs you a browser download and nothing per session
Your client spawns servers on demand and a cold start over twenty seconds per launch is unacceptable
You need file downloads, extensions or a persistent profile you control at the OS level
You want deterministic selector-driven scripts; act and observe route through a model, which is a second bill and a second source of flakiness

The bill for your context window

ToolEst. tokensDescribed?Required params
observe~79yes (39 chars)1
act~73yes (29 chars)1
navigate~70yes (17 chars)1
extract~65yes (26 chars)0
start~53yes (37 chars)0
end~52yes (37 chars)0

393 estimated tokens for the whole tools/list payload is cheap for what it buys, and the reason is that Browserbase pushed complexity out of the schema and into the model behind act, observe and extract. The six descriptions are terse: observe is the longest at 39 characters. That is a real tradeoff, not free savings. Your agent gets a small menu it can hold in working memory, and pays for the missing detail in retries when a natural-language action does not do what it read.

Setup reality

The npx package runs unauthenticated, which surprised us. When we booted it with a clean env it printed warnings that BROWSERBASE_API_KEY, BROWSERBASE_PROJECT_ID and MODEL_API_KEY were unset and substituted dummy values, then completed initialize anyway and served the full tool list. So all three come from your Browserbase account and your model provider, none are enforced at startup, and the failure surfaces later at session time. Transport is stdio. No resources are exposed. Most of that 21702 ms is npm fetching the package, so pin it locally if you launch often.

Questions people ask

Does the Browserbase MCP server need an API key to start?

No, and that is worth knowing before you debug. We started it with no environment set and it warned about three missing variables, used dummy values and served all six tools. Authentication failures appear when a session is actually created, not at boot.

How much context does it cost?

The tools/list response measured 1570 bytes, about 393 tokens using the chars/4 estimate, on 2026-08-19. That is far below the median of the servers in this batch and among the lightest browser servers we have measured.

What made the first launch drag?

Our run recorded 21702 ms from spawn to initialize, with npx resolving and downloading the package first. Install it into the project and the launch cost drops to node startup. The server itself reported protocol 2025-06-18 and version 3.0.0.

Is this the same as running Playwright locally?

No. The browser lives on Browserbase servers, so you get their IPs, captcha handling and session infrastructure, and they get every page your agent visits. Local Playwright servers cost nothing per session but give you none of that.

Other browsers servers we started

ServerVendorBootsContext tax
Playwright MCPMicrosoft~4,626 tokens
Puppeteer (archived)Anthropic (archived)~612 tokens

How this page is made: the server is spawned as an unprivileged user with a clean environment and no credentials, then asked for its tools, resources and prompts over stdio. Token figures are estimates at four characters per token, not a tokenizer count. One run, one machine. Corrections: contact the desk.

← All measured MCP servers