Hi Product Hunt, I am Nitai, I built Reflex. This started from my own annoyance. I was driving a browser from Claude Desktop, and every time the agent looked at a page it pulled in a huge accessibility tree and had to read all of it before it could do anything. On a heavy page that is tens of thousands of tokens for a single look, paid on every turn. Two or three pages in, my context was full, the run was slow, and the bill was real. So I built Reflex. It is a local MCP server that sends one compact view of the page, then only what changed after each action, and it can batch a whole task with its checks into one call instead of one round trip per click. In a benchmark where a real agent (Claude Opus 4.8) ran five live tasks through both tools, the Reflex runs finished about 3.5x faster end to end than Playwright MCP, about 24 seconds versus 83, on 2.3x fewer round trips and 7.7x less context. Both finished all five. On a Hacker News comment thread the gap is about 28x fewer tokens for the same look. And you can run browsers in parallel: give any call a session name and Reflex spins up a separate, isolated browser on demand, for free, so a swarm of subagents can each drive their own tab instead of fighting over one. The honest part: if your agent has a shell, the Playwright CLI and Vercel's agent browser are great and free, and on a flow you have already planned the CLI sends even fewer tokens than Reflex. Reflex is for clients that only speak MCP, like Claude Desktop, where those are out of reach. It is also not magic on every page; a list of 398 unique results does not compress, and there it pays about as much as Playwright. Everything runs on your machine, so your pages and logins never leave it. The free tier is 30 credits, about 10 to 15 tasks, no card. I would genuinely love to hear where it works for you and where it breaks, especially the pages where it still costs too much. I will be here all day to answer anything.

Forum Threads

p/reflex-mcp

•

7d ago

How much of your agent's context gets eaten just driving a browser?

Something I keep hitting with AI agents that browse: the browsing itself eats the budget. In an MCP client like Claude Desktop, every look at a page sends back the whole tree, so after two or three pages the context is full and the agent slows to a crawl. A simple task turns into a dozen round trips.

I want to hear how others handle this:

- If your agent drives a browser, how many pages before the context gets tight?

View all

Reflex MCP

A faster, parallel browser for your AI agent

A faster, parallel browser for your AI agent

Forum Threads

How much of your agent's context gets eaten just driving a browser?

Forum Threads

How much of your agent's context gets eaten just driving a browser?