Launching today
BetterWright
The ultra token-efficient browser runtime for AI agents
1 follower
The ultra token-efficient browser runtime for AI agents
1 follower
BetterWright is the ultra token-efficient, open-source browser runtime for AI agents. Its compressed snapshots use 50–90% fewer tokens than CDP and traditional browser-use setups, while a persistent profile keeps logins and state alive across turns. Agents get sandboxed Playwright, per-request network policy, encrypted credential filling without exposed secrets, proof screenshots, and CAPTCHA helpers. Use it from any shell-capable agent, MCP, Pi, or its Claude, Codex, and Grok OAuth loop.
Hey Product Hunt, I built BetterWright because browser agents spend far too much context just looking at pages, then lose the session or fall apart when a task reaches a login.
BetterWright turns the browser into a persistent runtime for agents. Its compressed accessibility snapshots use 50–90% fewer tokens than CDP and traditional browser-use setups, while refs and diff mode keep the next action precise. The same real profile survives across turns and days.
I also wanted it safe enough to run model-written code: Playwright is sandboxed, every request passes network policy, credentials are filled from an encrypted vault without entering the model context, and proof screenshots show what actually happened.
It works with shell-capable agents, MCP, and Pi, or through its own agent loop with Claude, Codex, or Grok OAuth.
It’s open source and MIT licensed. I’d really appreciate people trying real tasks and telling me where it wastes tokens or breaks.