llms.txt

# browsegrab
> Token-efficient browser agent for local LLMs — Playwright + AX tree + MarkGrab, MCP native.
`pip install browsegrab[all]`  (extras: cli, mcp, content)
- AX tree snapshots (~500-1.3K tokens/step) with short ref IDs (e1, e2, ...)
- Click/type/scroll/wait via ref IDs; agentic multi-step browse with loop detection
- 3 LLM providers (vLLM, Ollama, OpenAI-compat); MCP server (8 tools, stdio)
Exports: `BrowseGrabConfig`, `BrowseSession`, `BrowseResult`, `SnapshotResult`
```python
from browsegrab import BrowseGrabConfig, BrowseSession
async with BrowseSession(BrowseGrabConfig.from_env()) as s:
    result = await s.browse("https://example.com", "find pricing")
```
- [Full API](/llms-full.txt) | [Source](https://github.com/QuartzUnit/browsegrab) | [PyPI](https://pypi.org/project/browsegrab/)