Browser

The browser tools let an agent work in a real browser inside Synapse: open pages, read them, click, and fill in forms, while you watch and take over. Each chat and run gets its own browser.

The browser tools give an agent a Chromium that Synapse runs for it, one for each chat or run. You watch it live, take the wheel when you want, and press payments yourself. Built-in browser covers what that looks like, plus kept site sessions and recorded tasks.

A Browser agent holds every tool below, whatever else it holds. Any other agent can be given them one by one, as the Browser Automation tools.

Tools

Tool	Does
`browser_navigate`	Open an address
`browser_navigate_back`	Go back
`browser_snapshot`	Read the page: its text and the elements on it
`browser_find`	Find elements on the page
`browser_wait_for`	Wait for text to appear or go, or for a time
`browser_tabs`	List, open, close or switch tabs
`browser_hover`	Hover over an element
`browser_click`	Click an element
`browser_type`	Type into a field
`browser_fill_form`	Fill in several fields at once
`browser_select_option`	Choose in a dropdown
`browser_press_key`	Press a key
`browser_handle_dialog`	Accept or dismiss the page's own dialog
`browser_hand_over`	Give you the browser, saying what you should do
`browser_run_recording`	Repeat one of the agent's recordings by code

These are all it gets. Running the page's own JavaScript, reading files, the page's storage and network, and devtools are not offered, and a call to one is refused. The screenshots you see are taken by Synapse after each action, not by the agent.

What the agent reads comes back with the workspace's secrets scrubbed out.

What stops it

The browser holds every agent to the same rules, in code: Submit and Send only on your own words, payments and passwords always yours, and only public http or https addresses. A Browser agent set to Reads only sends nothing at all, and so does an agent that holds these tools without being a Browser agent.

Each browser also has limits: 40 actions for a task by default (a Browser agent's Actions per task, up to 200), at most 30 page loads a minute, and one load a second to each site.

Self-hosted

The Docker images include everything the browser needs. On an install from source, it needs:

Node, for npx: the browser's tools come from Playwright MCP, which Synapse starts itself.
Playwright's Chromium: playwright install chromium. Playwright keeps it under ~/.cache/ms-playwright; if yours is somewhere else, set playwright_browsers_path in settings.

The browser starts on an agent's first browser action, not with the server. Turning browser_automation_enabled off (it is on by default) removes the tools from the whole install.

Example requests

Open https://news.ycombinator.com and list the five top stories with their links.

Find order A-1001 on portal.saltmarsh.example and tell me its status.

Fill in the contact form on example.com with my name and email, and submit the form.

The second works signed in when a kept session for the site is picked for the agent, the chat or the step; the agent never signs in with a password. The third says submit the form, so the agent may press Submit. Without those words, it fills in the form and hands you the browser to press it.

Browser or Web Scraper

	Browser	Web Scraper
JavaScript rendering	Yes	Yes (crawl4ai)
You watch and can take over	Yes	No
Forms and signed-in sites	Yes, with you pressing what is yours	No
CAPTCHAs and bot checks	Handed to you	Stealth mode tries to avoid them
Many pages at once	No, one page at a time	Yes
Best for	Interactive sites, forms, sites you sign in to	Reading many pages quickly