Features
Each feature below is off unless you ask for it, apart from the core reading and learning.
Doing the task
| Feature |
What it does |
Details |
| Numbered page list |
The model sees each page as a short list of controls (#1 button "Add to cart"), not a screenshot |
What the agent can do |
| Learning per site |
Remembers click sequences that worked and replays them without a model call |
What the agent can do |
| Any model |
Anthropic, OpenAI, Ollama and other OpenAI-compatible servers, including local models |
Models |
| Hand-off to a person |
Pauses for a CAPTCHA, login or 2-step code, then carries on |
CAPTCHAs and hand-off |
| Pop-ups and permissions |
Sees JavaScript dialogs; browser permissions such as notifications are granted only when you say so |
Pop-ups and permissions |
Keeping things safe
| Feature |
What it does |
Details |
| Domain limits |
--allow-domain and --block-domain keep a run on the sites you chose |
Guardrails |
| Risky-click approval |
Asks before buy, pay, delete, send and sign-out clicks; blocks them when nobody can answer |
Guardrails |
| Injection watch |
Flags page text that tries to give the agent orders |
Guardrails |
| Budgets |
Stops a run past a token or time limit |
Guardrails |
| Personal data hidden from the model |
Emails, phones, cards and (with Presidio) names become <EMAIL_1>-style placeholders; the real value is typed back on the same site |
Personal data |
| Secrets |
Passwords are given by name and filled in at the keyboard; the model never sees them |
Passwords and secrets |
Seeing what happened
| Feature |
What it does |
Details |
| MCP server |
selfbrowse-mcp gives Claude, VS Code and other MCP clients a browser, as one task tool or step by step |
MCP server |
| Python API |
BrowseAgent(...), custom guardrails, step callbacks |
Get started |