Built-in browser

Each chat and each run that uses the browser gets its own: a Chromium that Synapse runs for it, which you watch live. The agent reads and clicks through pages; you take the wheel whenever you want, and some things are always yours to do: passwords, card details, codes and payments.

An agent works in the browser when it is a Browser agent, or holds the browser tools. The site sessions and recordings below are under Agents › Browser, on its two tabs, Sessions and Recordings.

Watching it work

In a chat, the browser opens in a pane beside the conversation when the agent first uses it. A strip under the address says who is driving: the agent, you, or nobody. Under the page, What the agent did lists each action. Closing the pane only stops you watching; the agent carries on.

On a run's page, a running browser shows under Live browser, at the page's own shape. Enlarge opens it large. After the run, The browser shows a screenshot of each action the agent took.

Stop, at the top of the pane, stops the agent and closes its browser. In a chat it ends the agent's turn. On a run's page it cancels the whole run, not only the step.

A screenshot is kept after each agent action, with the run or the chat, so you can look back once the browser has closed. Nothing is kept while you drive. The screenshots, and the record of who pressed and sent what, go when the run's log or the chat is deleted.

A browser nobody is watching closes after 10 minutes without an action.

Taking over

Press Take over to use the page yourself. If you click the page while the agent is driving, Synapse asks first: Keep watching or Take over. The agent finishes the action it is doing, then waits.

While you drive, your clicks and typing reach the page, and the address bar takes an address. When you're done:

1. Optionally, type a note for the agent in the strip, such as I signed in, carry on.
2. Press Hand back.

The agent is told how long you drove, your note, and the page it is now on, and reads the page before it acts again. A note that holds a card number is refused: card details go in the page, never to the agent.

The agent hands you the browser itself when a step is yours. The chat or the run shows Your turn in the browser, with Open the browser. That happens when:

• a field is a password, a card number or a one-time code;
• a CAPTCHA or bot check is showing;
• a button pays, deletes something for good or closes an account;
• a button sends something for you and you haven't asked for it (see What it may press);
• the agent wants to type on a site that isn't one of its sites, and the strip asks you to allow it.

When you hand back, a waiting run resumes its step and a chat carries on with your note as its next message.

Keeping agents signed in

Synapse can keep a site's signed-in session for your agents: the site's cookies and storage, never your password. Agents never see or type a password.

1. Open Agents › Browser, on the Sessions tab.
2. Press Keep me signed in to a site, type the site's address, and press Open the browser.
3. Sign in, in the browser that opens. It is yours alone: no agent drives it and nothing is recorded.
4. Press Keep.

A session lasts until the site ends it, and never more than 30 days. Its row says when it expires. An expired session keeps its place in every agent and step that picked it: press Sign in again on its row. Remove deletes it, and says first what picks it.

You pick sessions in three places, one per site:

• a Browser agent's Browser section, beside each of its sites;
• a chat, with Sessions at the top of the browser pane;
• a workflow step's Sessions for this step.

A browser starts with all of them. Where two places pick a session for the same site, the nearer one wins: the step's, then the chat's, then the agent's. Synapse matches each session to its site, and the agent never chooses one.

A session is as good as a password while it lasts, so Synapse keeps it encrypted and opens it only for the person who kept it (on Cloud, also for the people they share it with).

Recording a task

Do a task once while Synapse records it, and agents and workflows can repeat it.

1. Open Agents › Browser, on the Recordings tab, and press Record a task.
2. Optionally, pick Sessions it starts with and a Start at page.
3. Press Open the browser, and do the task as you normally would. Each thing you do is a step: open an address, click, type, choose, press Enter, go back.
4. Press Stop and keep, or Discard.

What you type into a password, card or one-time-code field is never kept, and neither is anything that looks like a card number. That step stays yours: when the recording is repeated, it hands you the browser there.

The recording's page lists its steps, each with a picture of the page after it. You can rename a step, delete it, or Re-record from here: the steps before it replay, then you do it again from that step. Values you typed can become inputs, so each run types its own (Make it an input on a step). Every saved change is a version, under Versions.

Repeating a recording

A recording is repeated by code, step by step, with no model calls while the pages still match. Each step finds its target the way you'd describe it: by its role and name first, then its label, then its text, and its position on the page last.

• Presses that send something for you (submit, send, post, book…) run in a chat, because you pressed them when recording. In a workflow run, which nobody watches, only the steps you tick under Sends on a Browser step as may send on its own do; at any other, the run waits for you.
• Payments, permanent deletes and closing an account are never replayed. The browser is handed to you there, every time.

When a page has changed

A step that can't find its target, or doesn't reach the page it expects, stops the replay there. What happens next is set on the Browser step, under If a page changed:

• Pause for you (the default): the run waits for you to do that step in its browser.
• Let a Browser agent find the way: the agent you pick does that step for this run.

Either way, the way found is kept as a fix for the recording, never applied by itself. It waits on the recording's page under Fixes waiting: Approve fix makes it the recording's next version, and Discard fix drops it. Nothing changes until it is approved. A step you re-record waits for your approval the same way.

A Browser step or a Browser agent

A Browser step repeats a recording in the run's own browser. In the step, pick the Recording, fill its Inputs from the run ({{ state.key }} or text), tick which Sends may go on their own, and choose what happens If a page changed. What the recording reads back is kept under the step's output.

A Browser agent is an agent whose type is Browser. Its Browser section sets what no page can change:

• How it works the page. Reads only, the default, can read and click through pages, and nothing it does sends anything: forms, orders and messages are stopped in the browser. Works with you can fill in and submit forms, within the rules below.
• Sites it may use. Listed, it opens no page anywhere else, even when a page links there, and starts signed in to each site you pick a session for. With none listed, it may open any public site, and asks you before typing on one.
• Actions per task, 40 by default, 1 to 200. At the limit it stops and says so.

A Browser agent can also run a recording as one of its actions, when it has one. Ask Synapse AI to give it one. Recordings are made by a person, in the app; Synapse AI never records one.

What it may press

These rules are code in the browser itself, not instructions to the model, so no page can talk an agent past them:

• Submit, Send and the like only when you asked for it in your own words. A press that would send, post, book or submit a form is stopped, and the agent must quote your request back. Without a quote, it hands you the browser to press it yourself.
• Payments are always yours to press: a payment form, a checkout, or a button that pays, orders or subscribes. So are permanent deletes and closing an account. What the agent sends to a payment provider or a checkout is held for you.
• Passwords, card details and one-time codes are always yours to type. The agent stops at the field and hands you the browser.
• Public addresses only. Every request a page makes, redirects and in-page scripts included, goes only to a public http or https address: never to your own machine or private network.
• Reads only means nothing sends. In a Reads only browser every request that would change something is stopped, whatever the agent presses.

The pane says it once: agents can make mistakes, and you always press payments yourself.

Self-hosted and Cloud

The browser works the same way in both. The differences:

Self-hosted, the Docker images include everything the browser needs. On an install from source, it needs Node (for npx) and Playwright's Chromium: playwright install chromium. The browser starts on the first browser action; if it can't, the error says Synapse's browser did not start. Is Chromium installed (playwright install chromium)?

On SynapseOrch Cloud, every plan has the browser in a chat. A browser in a workflow run (scheduled ones included), sessions and recordings come with Solo and up. Each plan also sets how many browsers run at once, and how long one may run. At launch:

A browser waiting for you doesn't count as running. Your workspace's own limits are under Settings → Plan.