Introduction

WhittleBrowser lets an AI agent use a web browser while sending the model far fewer tokens, and keeping secrets and hidden prompt injections away from it. It runs as an MCP server, so Claude Code, Cursor, Codex, VS Code and any other MCP client can use it, and it is also a TypeScript library. It is free and MIT-licensed.

What it does

• Cheap to read. A page becomes compact, ref-stable text within a token budget. After an action the model gets only what changed. On our 11-page benchmark one page view costs about 96% fewer tokens than Playwright MCP. In real agent runs on whole tasks the saving is smaller and depends on the task: large or re-rendering pages save the most, small pages save little. See the benchmark for every number and how it was measured.
• Safe by default. Secrets are masked and hidden prompt injections are dropped before text reaches the model. Sensitive pages (banks, payment forms, credential fields) pause the agent. Irreversible actions need approval.
• Watchable. X-ray overlays what the agent sees, and an optional ghost cursor shows what it is about to do.
• Fast at research. Read several pages in parallel and get one compact digest.
• Honest. It identifies itself and hands CAPTCHAs to a person instead of trying to defeat bot protection.

What's available

• Engine (this documentation): open source and available now. It is the MCP server and library described on every page here.
• Cloud: the engine hosted behind an API. Not open yet; see Cloud API and usage and the waitlist.
• WhittleBrowser for Mac: a desktop browser built on the engine. Coming soon.

Where to next

• Getting started: install and run it. The package is @whittlebrowser/engine on npm.
• Use with Claude Code or other AI tools: connect an agent in one command.
• How it works: the in-page agent, refs, serializer, diffs.
• Protocol: the exact view format and every tool.
• Safety model: each threat and its defence.
• Benchmark and comparison.