Forensic failures
Near-miss candidates, console errors, and a screenshot path — enough to repair in one agent turn.
Local browser automation for MCP clients
Browser automation that doesn't eat your context window
Reflex runs locally in Claude Desktop, Cursor, Claude Code, or any MCP client. Batch a whole flow into about 2 guarded calls per task (6 across three measured flows vs 14 with Playwright MCP), send ~13K tokens into context instead of ~56K, and never let a page leave your machine.
For AI assistants with browser tools: research products, fill forms, check prices, scrape tables, and compare options without burning your context on page dumps.
- Fewer round trips
- ~4× fewer tokens per task
- Private by design
Get 30 free credits
Read the benchmarkNo card. 30 free credits (~~10-15 tasks). 2 minutes to first task.
Wikipedia: search an article, open its revision history
Playwright MCP5/5 calls
tokens into context~30K
Reflex2/2 calls
tokens into context~8K
Drop Reflex into your client in one click.
One click in your client, then reflex_auth to sign in.
Getting started
Two minutes to your first task.
your MCP client
# 1 — install Reflex
one click into Cursor or VS Code, or npx -y reflex-browser mcp in your MCP config
# 2 — sign in once
ask your agent to run reflex_auth { "action": "login" } — opens your browser, saves a key locally. 30 free credits, no card
# 3 — ask your agent, in plain English
> Compare the prices of these three products and tell me which has free shipping.
Reflex batches the whole flow into one guarded call and reports back what changed, not the whole page.
Benchmarks
Measured, not promised.
Numbers from our published head-to-head, run 2026-06-12 against @playwright/mcp 0.0.76 on live sites, no LLM on either side. Every number, and how to reproduce it, is on the benchmarks page.
| Playwright MCP | Reflex | |
|---|---|---|
| Tool calls, 3 measured flows | 14 | 6 |
| Tokens into context, same flows | ~56K | ~13K |
| Flows completed, of 6 attempted | 3 | 5 |
| One look at a Wikipedia article | ~31K tokens | ~4K tokens |
| One look at the worst page tested (W3C CSS spec) | ~169K tokens | ~7K tokens |
| Re-reading an unchanged page | ~16K tokens | 91 characters: "(page unchanged)" |
Measured 2026-06-12. See every number and how to reproduce it.
Pure MCP clients like Claude Desktop pay the inline snapshot price above on every look. Reflex distills the page instead, so your agent keeps its context window for the actual work. Playwright MCP is free and better for iframes, network mocking, and Firefox/WebKit; Reflex wins when your client pays per snapshot in context. Full methodology and the reproduce command are on the benchmarks page.
Agent-native
Built for agents, not just browsers.
Batched verification
Expectations checked inside a single guarded call, so a flow lands or fails loudly without extra round trips.
0-token replay
Saved flows re-run with no model in the loop and self-heal when the page shifts.
Watch the real run
See it in action.
Privacy
Your pages never leave your machine.
Reflex runs entirely on your computer: the MCP server, the browser, your logins, every page you visit. The only things that ever cross the network are your API key and a count of tool calls, so we can meter credits. We could not read your pages, sessions, or passwords if we wanted to; the architecture has no path for them.
Compare that to cloud browser products, where every page you visit renders on someone else's server.
Your machine stays here
Claude Desktop
or Cursor, Claude Code, any MCP client
Reflex MCP server
runs locally
Your Chrome
your real browser and logins
Pages, sessions, passwords
read and acted on locally
API key + call count
pages never cross
Reflex servers
- Is this key valid?
- Credit balance
- Usage counts
- Paddle payments
No path exists for your pages, logins, or sessions to reach us.
5/6 vs 3/6
flows completed on the same 6 live-site tasks (first-try success)
7.9x fewer tokens
per page look vs Playwright MCP, averaged across 12 live sites
0 tokens
to re-run a saved flow (deterministic replay, no model in the loop)
Pricing
Pay for what you use. Paid credits never expire.
Full pricing details →Starter
$9one-time
For occasional flows.
- 150 credits
- ~50-75 tasks
- $0.12-0.18 per task
Most popular
Pro
$19one-time
Best value for steady use.
- 500 credits
- ~165-250 tasks
- $0.08-0.11 per task
Power
$49one-time
For heavy, daily automation.
- 1,800 credits
- ~600-900 tasks
- $0.05-0.08 per task
FAQ
Questions, answered straight.
Everything runs locally: the MCP server, your Chrome, your logins, the pages. The only data that reaches our servers is your email (for your account), your API key, and a counter of tool calls for credit metering. Page content, URLs you visit, credentials, and session cookies never leave your machine; there is no code path that could send them. Payments are handled by Paddle, our merchant of record; we never see card numbers. Our privacy policy is short because there is almost nothing to disclose.
Because browser automation is bursty. Most people do not use an agent browser every day, and a subscription punishes that. You buy a pack, paid credits never expire, and a typical task costs 2-3 credits. No renewal anxiety, no cancel-flow dark patterns.
Nothing breaks. The next tool call returns a polite message with a top-up link instead of running, and Claude relays it to you. Your installation, saved flows, and login states are untouched; buy a pack and the same call works again seconds later. We email you once when you hit zero, and that's it.
Reflex works with any client that can run local MCP servers: Claude Desktop, Claude Code, Cursor, Windsurf, VS Code, and others. ChatGPT currently supports remote connectors rather than local MCP servers, so not yet; the moment that changes, Reflex will work there too, since it speaks standard MCP.
It is, and it's good; Microsoft maintains it and it wins at iframes, network mocking, and Firefox/WebKit. The difference is the cost model inside your chat: in our measured head-to-head the same flows took Playwright MCP 14 tool calls and ~56K tokens of your context window versus 6 calls and ~13K for Reflex, and it completed 3 of 6 flows versus Reflex's 5. Every round trip is seconds of model time, and every token of page dump crowds out your actual conversation. We published the full benchmark, including the places Reflex loses, at /benchmarks. If those numbers don't matter to your usage, use Playwright MCP; it's a fine tool.
Reflex drives your installed Chrome with its own profile. Log in to a site once in that window and tell Reflex to save the session; it stores cookies and localStorage locally on your disk and reuses them on demand. Your logins persist across tasks and never leave your machine.
One credit per tool call, and a typical task is 2-3 calls: one to open and read the page, one batched call that performs every step and verifies the result. That's the core design difference; tools that take one action per call need 5-6 calls per flow for the same work (14 vs 6 across our three measured flows). Failed calls also cost one credit, because the failure report you get back (what missed, the closest candidates, console errors, a screenshot) is usually enough to fix the task in one more turn.
Only for the metering layer, and it never blocks your work. The key is validated when the server starts, credits are decremented against a local cached balance, and usage syncs in the background with an offline buffer. Browsing localhost or an intranet app works fine.
Paid credits never expire. The 30 free signup credits expire after 14 days; they exist so you can evaluate Reflex properly, not as a permanent free tier.
If you bought a pack and it isn't working out, email support within 14 days and we refund the unused balance, no interrogation. Refunds for consumed credits are case by case, biased toward yes; arguing over $9 is not a business we want to be in.
Ready when you are
Same browser tasks. ~13K tokens, not ~56K.
Installed in 2 minutes. Your pages stay yours.