# Fantoma

> A browser for AI agents that reads pages through the accessibility tree instead of screenshots. Python library, HTTP server and MCP server. MIT licence.

Reading pages works on small local models (30/31 sites on a 35B local MoE). Multi-step interaction needs a stronger model: the same local model finished 1 of 6 flows, a frontier model finished 6 of 6. Trust `verified` over `success` on results.

## Use it
- Install: `pip install "fantoma[mcp]"`, then `claude mcp add fantoma -- fantoma-mcp`
- MCP tools: `fantoma_read` and `fantoma_login` (no LLM needed), `fantoma_run` and `fantoma_extract` (need an OpenAI-compatible LLM)

## Docs
- [README](https://github.com/Huzy85/fantoma/blob/main/README.md): install, MCP setup, safety, limitations, measured results
- [Changelog](https://github.com/Huzy85/fantoma/blob/main/CHANGELOG.md)

## Limits
- No automation is undetectable. Protected-site reads are good, not guaranteed.
- For accessibility auditing, QA of your own apps and personal automation. No bulk account creation.
