What it does
Browsegrab is a browser automation library that combines Playwright's accessibility tree with MarkGrab's HTML-to-markdown conversion. Elements receive stable references such as e1 and e2, so the model does not need a vision model to act on them. It also caches successful workflows, so repeated workflows can run with zero LLM calls, and it includes a 5-stage JSON parser for action outputs. Its built-in MCP server exposes eight tools: navigate, click, type, snapshot, scroll, extract content, go back and wait.
It is intended for agents run by local models, such as those served through vLLM, Ollama or OpenAI-compatible endpoints, where the tokens sent per step matter. The publisher reports about 500 to 1,500 tokens per step, compared with 4,000 to 10,000 for browser-use, and about 60 tokens for a single-element page such as example.com versus about 1,254 for a Wikipedia article with 452 interactive elements.
It can be used through a Python API, a CLI or as an MCP server connected to clients such as Claude Desktop, Cursor or VS Code. Settings are read from environment variables with the BROWSEGRAB_ prefix. It is part of the QuartzUnit ecosystem alongside markgrab, snapgrab and docpick, and is released under the MIT license.
Limitations and requirements
Designed for local LLMs of roughly 8B to 35B parameters. Core dependencies are playwright and httpx. Configuration is done through BROWSEGRAB_* environment variables.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
Browsegrab
Token-efficient browser agent for local LLMs. Playwright + accessibility tree + MarkGrab.
v0.1.1As shown on registry.modelcontextprotocol.io when we last read it