Chromote: Headless Chrome Remote Interface

Chromote is an R implementation of the Chrome DevTools Protocol. It works with Chrome, Chromium, Opera, Vivaldi, and other browsers based on Chromium. By default it uses Google Chrome (which must already be installed on the system). To use a different browser, see Specifying which browser to use.

Chromote is not the only R package that implements the Chrome DevTools Protocol. Here are some others:

crrri by Romain Lesur and Christophe Dervieux
decapitated by Bob Rudis
chradle by Miles McBain

The interface to Chromote is similar to chrome-remote-interface for node.js.

Installation

# CRAN
install.packages("chromote")

# Development
remotes::install_github("rstudio/chromote")

Basic usage

This will start a headless browser and open an interactive viewer for it in a normal browser, so that you can see what the headless browser is doing.

library(chromote)

b <- ChromoteSession$new()

# In a web browser, open a viewer for the headless browser. Works best with
# Chromium-based browsers.
b$view()

The browser can be given commands, as specified by the Chrome DevTools Protocol. For example, $Browser$getVersion() (which corresponds to the Browser.getVersion in the API docs) will query the browser for version information:

b$Browser$getVersion()
#> $protocolVersion
#> [1] "1.3"
#>
#> $product
#> [1] "HeadlessChrome/98.0.4758.102"
#>
#> $revision
#> [1] "@273bf7ac8c909cde36982d27f66f3c70846a3718"
#>
#> $userAgent
#> [1] "Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) HeadlessChrome/98.0.4758.102 Safari/537.36"
#>
#> $jsVersion
#> [1] "9.8.177.11"

If you have the viewer open and run the following, you’ll see the web page load in the viewer:

b$Page$navigate("https://www.r-project.org/")

In the official Chrome DevTools Protocol (CDP) documentation, this is the Page.navigate command.

In addition to full support of the CDP, ChromoteSession objects also some convenience methods, like $screenshot(). (See the Examples section below for more information about screenshots.)

# Saves to screenshot.png
b$screenshot()

# Takes a screenshot of elements picked out by CSS selector
b$screenshot("sidebar.png", selector = ".sidebar")

Note: All members of Chromote and ChromoteSession objects which start with a capital letter (like b$Page, b$DOM, and b$Browser) correspond to domains from the Chrome DevTools Protocol, and are documented in the official CDP site. All members which start with a lower-case letter (like b$screenshot and b$close) are not part of the Chrome DevTools Protocol, and are specific to Chromote and ChromoteSession.

Here is an example of how to use Chromote to find the position of a DOM element using DOM.getBoxModel.

x <- b$DOM$getDocument()
x <- b$DOM$querySelector(x$root$nodeId, ".sidebar")
x <- b$DOM$getBoxModel(x$nodeId)
str(x)
#> List of 1
#>  $ model:List of 6
#>   ..$ content:List of 8
#>   .. ..$ : num 128
#>   .. ..$ : int 28
#>   .. ..$ : num 292
#>   .. ..$ : int 28
#>   .. ..$ : num 292
#>   .. ..$ : num 988
#>   .. ..$ : num 128
#>   .. ..$ : num 988
#>   ..$ padding:List of 8
#>   .. ..$ : num 112
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : num 988
#>   .. ..$ : num 112
#>   .. ..$ : num 988
#>   ..$ border :List of 8
#>   .. ..$ : num 112
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : num 988
#>   .. ..$ : num 112
#>   .. ..$ : num 988
#>   ..$ margin :List of 8
#>   .. ..$ : int 15
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : num 1030
#>   .. ..$ : int 15
#>   .. ..$ : num 1030
#>   ..$ width  : int 195
#>   ..$ height : int 960

Or you can do the same thing by chaining commands together with a magrittr pipe:

b$DOM$getDocument() %>%
  { b$DOM$querySelector(.$root$nodeId, ".sidebar") } %>%
  { b$DOM$getBoxModel(.$nodeId) } %>%
  str()
#> List of 1
#>  $ model:List of 6
#>   ..$ content:List of 8
#>   .. ..$ : num 128
#>   .. ..$ : int 28
#>   .. ..$ : num 292
#>   .. ..$ : int 28
#>   .. ..$ : num 292
#>   .. ..$ : num 988
#>   .. ..$ : num 128
#>   .. ..$ : num 988
#>   ..$ padding:List of 8
#>   .. ..$ : num 112
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : num 988
#>   .. ..$ : num 112
#>   .. ..$ : num 988
#>   ..$ border :List of 8
#>   .. ..$ : num 112
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : num 988
#>   .. ..$ : num 112
#>   .. ..$ : num 988
#>   ..$ margin :List of 8
#>   .. ..$ : int 15
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : num 1030
#>   .. ..$ : int 15
#>   .. ..$ : num 1030
#>   ..$ width  : int 195
#>   ..$ height : int 960

Creating new tabs and managing the process

To create a new tab/window:

b1 <- b$new_session()

Once it’s created, you can perform operations with the new tab without affecting the first one.

b1$view()
b1$Page$navigate("https://github.com/rstudio/chromote")
#> $frameId
#> [1] "714439EBDD663E597658503C86F77B0B"
#>
#> $loaderId
#> [1] "F39339CBA7D1ACB83618FEF40C3C7467"

To close a browser tab/window, you can run:

b1$close()

This is different from shutting down the browser process. If you call b$close(), the browser process will still be running, even if all tabs have been closed. If all tabs have been closed, you can still create a new tab by calling b1$new_session().

To shut down the process, call:

b1$parent$close()

b1$parent is a Chromote object (as opposed to ChromoteSession), which represents the browser as a whole. This is explained in The Chromote object model.

Commands and Events

The Chrome DevTools Protocol has two types of methods: commands and events. The methods used in the previous examples are commands. That is, they tell the browser to do something; the browser does it, and then sends back some data.

Events are quite different from commands. When, for example, you run b$Page$loadEventFired(), it does not send a message to the browser. Rather, this method tells the R process to wait until it receives a Page.loadEventFired message from the browser.

Here is an example of how that event can be used. Note that these two lines of code must be run together, without any delay at all (this can be enforced by wrapping both lines of code in { .... }).

# Send a command to navigate to a page
b$Page$navigate("https://www.r-project.org")
#> $frameId
#> [1] "0ADE3CFBAF764B0308ADE1ACCC33358B"
#>
#> $loaderId
#> [1] "112AF4AC0C13FF4A95BED8173C3F4C7F"

# Wait for the Page.loadEventFired event
b$Page$loadEventFired()
#> $timestamp
#> [1] 680.7603

After running these two lines, the R process will be blocked. While it’s blocked, the browser will load the page, and then send a message to the R process saying that the Page.loadEventFired event has occurred. The message looks something like this:

{"method":"Page.loadEventFired","params":{"timestamp":699232.345338}}

After the R process receives this message, the function returns the value, which looks like this:

$timestamp
[1] 699232.3

Note: This sequence of commands, with Page$navigate() and then Page$loadEventFired() will not work 100% of the time. See Loading a Page Reliably for more information.

Technical note: Chromote insulates the user from some of the details of how the CDP implements event notifications. Event notifications are not sent from the browser to the R process by default; you must first send a command to enable event notifications for a domain. For example Page.enable enables event notifications for the Page domain – the browser will send messages for all Page events. (See the Events section in this page). These notifications will continue to be sent until the browser receives a Page.disable command.

By default, Chromote hides this implementation detail. When you call b$Page$loadEventFired(), Chromote sends a Page.enable command automatically, and then waits until it receives the Page.loadEventFired event notification. Then it sends a Page.disable command.

Note that in asynchronous mode, the behavior is slightly more sophisticated: it maintains a counter of how many outstanding events it is waiting for in a given domain. When that count goes from 0 to 1, it sends the X.enable command; when the count goes from 1 to 0, it sends the X.disable command. For more information, see the Async events section.

If you do not want automatic event enabling and disabling, then when creating the ChromoteSession object, use ChromoteSession$new(auto_events = FALSE).

The Chromote object model

There are two R6 classes that are used to represent the Chrome browser. One is Chromote, and the other is ChromoteSession. A Chromote object represents the browser as a whole, and it can have multiple targets, which each represent a browser tab. In the Chrome DevTools Protocol, each target can have one or more debugging sessions to control it. A ChromoteSession object represents a single session.

When a ChromoteSession object is instantiated, a target is created, then a session is attached to that target, and the ChromoteSession object represents the session. (It is possible, though not very useful, to have multiple ChromoteSession objects connected to the same target, each with a different session.)

A Chromote object can have any number of ChromoteSession objects as children. It is not necessary to create a Chromote object manually. You can simply call:

b <- ChromoteSession$new()

and it will automatically create a Chromote object if one has not already been created. The Chromote package will then designate that Chromote object as the default Chromote object for the package, so that any future calls to ChromoteSession$new() will automatically use the same Chromote. This is so that it doesn’t start a new browser for every ChromoteSession object that is created.

In the Chrome DevTools Protocol, most commands can be sent to individual sessions using the ChromoteSession object, but there are some commands which can only be sent to the overall browser, using the Chromote object.

To access the parent Chromote object from a ChromoteSession, you can simply use $parent:

b <- ChromoteSession$new()
m <- b$parent

With a Chromote object, you can get a list containing all the ChromoteSessions, with $get_sessions():

m$get_sessions()

Normally, subsequent calls to ChromoteSession$new() will use the existing Chromote object. However, if you want to start a new browser process, you can manually create a Chromote object, then spawn a session from it; or you can pass the new Chromote object to ChromoteSession$new():

cm <- Chromote$new()
b1 <- cm$new_session()

# Or:
b1 <- ChromoteSession$new(parent = cm)

Note that if you use either of these methods, the new Chromote object will not be set as the default that is used by future calls to ChromoteSesssion$new(). See Specifying which browser to use for information on setting the default Chromote object.

There are also the following classes which represent the browser at a lower level:

Browser: This represents an instance of a browser that supports the Chrome DevTools Protocol. It contains information about how to communicate with the Chrome browser. A Chromote object contains one of these.
Chrome: This is a subclass of Browser that represents a local browser. It extends the Browser class with a processx::process object, which represents the browser’s system process.
ChromeRemote: This is a subclass of Browser that represents a browser running on a remote system.

Debugging

Calling b$debug_messages(TRUE) will enable the printing of all the JSON messages sent between R and Chrome. This can be very helpful for understanding how the Chrome DevTools Protocol works.

b <- ChromoteSession$new()
b$parent$debug_messages(TRUE)
b$Page$navigate("https://www.r-project.org/")
#> SEND {"method":"Page.navigate","params":{"url":"https://www.r-project.org/"},"id":53,"sessionId":"12CB6B044A379DA0BDCFBBA55318247C"}
#> $frameId
#> [1] "BAAC175C67E55886207BADE1776E7B1F"
#>
#> $loaderId
#> [1] "66DED3DF9403DA4A307444765FDE828E"

# Disable debug messages
b$parent$debug_messages(FALSE)

Synchronous vs. asynchronous usage

By default, when you call methods from a Chromote or ChromoteSession object, it operates in synchronous mode. For example, when you call a command function (like b$Page$navigate()), a command message is sent to the headless browser, the headless browser executes that command, and it sends a response message back. When the R process receives the response, it converts it from JSON to an R object and the function returns that value. During this time, the R process is blocked; no other R code can execute.

The methods in Chromote/ChromoteSession objects can also be called in asynchronous mode. In async mode, a command function fires off a message to the browser, and then the R process continues running other code; when the response comes back at some time in the future, the R process calls another function and passes the response value to it.

There are two different ways of using async with Chromote. The first is with promises (note that these are not the regular R-language promises; these are similar to JavaScript promises for async programming.) The second way is with callbacks: you call methods with a callback_ argument. Although callbacks are initially easier to use than promises, once you start writing more complex code, managing callbacks becomes very difficult, especially when error handling is involved. For this reason, this document will focus mostly on promises instead of callback-style programming.

When Chromote methods are called in synchronous mode, under the hood, they are implemented with asynchronous functions, and then waiting for the asynchronous functions to resolve.

Technical note about the event loop: When methods are called asynchronously, the R process will run callbacks and promises using an event loop provided by the later package. This event loop is very similar to the one used in JavaScript, which is explained in depth by Philip Roberts in this video. One important difference between JavaScript’s event loop and the one provided by later’s is that in JavaScript, the event loop only runs when the call stack is empty (essentially, when the JS runtime is idle); with later the event loop similarly runs when the call stack is empty (when the R console is idle), but it can also be run at any point by calling later::run_now().

There is another important difference between the JS event loop and the one used by Chromote: Chromote uses private event loops provided by later. Running the private event loop with run_now() will not interfere with the global event loop. This is crucial for being able to run asynchronous code in a way that appears synchronous.

Why async?

The synchronous API is easier to use than the asynchronous one. So why would you want to use the async API? Here are some reasons:

The async API allows you to send commands to the browser that may take some time for the browser to complete, and they will not block the R process from doing other work while the browser executes the command.
The async API lets you send commands to multiple browser “tabs” and let them work in parallel.

On the other hand, async programming can make it difficult to write code that proceeds in a straightforward, linear manner. Async programming may be difficult to use in, say, an analysis script.

When using Chromote interactively at the R console, it’s usually best to just call methods synchronously. This fits well with a iterative, interactive data analysis workflow.

When you are programming with Chromote instead of using it interactively, it is in many cases better to call the methods asynchronously, because it allows for better performance. In a later section, we’ll see how to write asynchronous code with Chromote that can be run either synchronously or asynchronously. This provides the best of both worlds.

Async commands

When a method is called in synchronous mode, it blocks until the browser sends back a response, and then it returns the value, converted from JSON to an R object. For example:

# Synchronous
str(b$Browser$getVersion())
#> List of 5
#>  $ protocolVersion: chr "1.3"
#>  $ product        : chr "HeadlessChrome/98.0.4758.102"
#>  $ revision       : chr "@273bf7ac8c909cde36982d27f66f3c70846a3718"
#>  $ userAgent      : chr "Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) HeadlessChrome/98.0.4758.102 Safari/537.36"
#>  $ jsVersion      : chr "9.8.177.11"

In async mode, there are two ways to use the value that the browser sends to the R process. One is to use the callback_ argument with wait_=FALSE. The wait_=FALSE tells it to run the command in async mode; instead of returning the value from the browser, it returns a promise. For example:

# Async with callback
b$Browser$getVersion(wait_ = FALSE, callback_ = str)
#> <Promise [pending]>
#> List of 5
#>  $ protocolVersion: chr "1.3"
#>  $ product        : chr "HeadlessChrome/98.0.4758.102"
#>  $ revision       : chr "@273bf7ac8c909cde36982d27f66f3c70846a3718"
#>  $ userAgent      : chr "Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) HeadlessChrome/98.0.4758.102 Safari/537.36"
#>  $ jsVersion      : chr "9.8.177.11"

Notice that the function returned <Promise [pending]>, and then it printed out the data. We’ll come back to the promise part.

Technical note: When you pass a function as callback_, that function is used as the first step in the promise chain that is returned.

If you run the command in a code block (or a function), the entire code block will finish executing before the callback can be executed. For example:

{
  b$Browser$getVersion(wait_ = FALSE, callback_ = str)
  1+1
}
#> [1] 2
#> List of 5
#>  $ protocolVersion: chr "1.3"
#>  $ product        : chr "HeadlessChrome/98.0.4758.102"
#>  $ revision       : chr "@273bf7ac8c909cde36982d27f66f3c70846a3718"
#>  $ userAgent      : chr "Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) HeadlessChrome/98.0.4758.102 Safari/537.36"
#>  $ jsVersion      : chr "9.8.177.11"

In the code above, it executes the 1+1 and returns the value before the str callback can be executed on the message from the browser.

If you want to store the value from the browser, you can write a callback that stores the value like so:

# This will extract the product field
product <- NULL
b$Browser$getVersion(wait_ = FALSE, callback_ = function(msg) {
  product <<- msg$product
})
#> <Promise [pending]>
# Wait for a moment, then run:
product
#> [1] "HeadlessChrome/98.0.4758.102"

But to get the value, you need to wait for the callback to execute before you can use the value. Waiting for a value is simple when running R interactively – you can just add a message("message arrived") call in the callback and wait for it before running the next line of code – but waiting for the value is not easy to do using ordinary straight-line coding. Fortunately, Chromote has a way to wait for async operations, which we’ll see later.

The other way of using the value is to use promises. If wait_=FALSE and no callback_ is passed to the command, then it will simply return a promise. Promises have many advantages over plain old callbacks: they are easier to chain, and they provide better error-handling capabilities. You can chain more steps to the promise: when the promise resolves – that is, when the message is received from the browser – it will run the next step in the promise chain.

Here’s an example that uses promises to print out the version information. Note that the surrounding curly braces are there to indicate that this whole thing must be run as a block without any idle time in between the function calls – if you were to run the code in the R console line-by-line, the browser would send back the message and the promise would resolve before you called p$then(), which is where you tell the promise what to do with the return value. (The curly braces aren’t strictly necessary – you could run the code inside the braces in a single paste operation and have the same effect.)

{
  p <- b$Browser$getVersion(wait_ = FALSE)
  p$then(function(value) {
    print(value$product)
  })
}
# Wait for a moment, then prints:
#> [1] "HeadlessChrome/98.0.4758.102"

Here are some progressively more concise ways of achieving the same thing. As you work with promises, you will see these various forms of promise chaining. For more information, see the promises documentation.

library(promises)

# Regular function pipe to then()
b$Browser$getVersion(wait_ = FALSE) %>% then(function(value) {
  print(value$product)
})

# Promise-pipe to anonymous function, which must be wrapped in parens
b$Browser$getVersion(wait_ = FALSE) %...>% (function(value) {
  print(value$product)
})

# Promise-pipe to an expression (which gets converted to a function with the first argument `.`)
b$Browser$getVersion(wait_ = FALSE) %...>% { print(.$product) }

# Promise-pipe to a named function, with parentheses
print_product <- function(msg) print(msg$product)
b$Browser$getVersion(wait_ = FALSE) %...>% print_product()

# Promise-pipe to a named function, without parentheses
b$Browser$getVersion(wait_ = FALSE) %...>% print_product

The earlier example where we found the dimensions of a DOM element using CSS selectors was done with the synchronous API and %>% pipes. The same can be done in async mode by switching from the regular pipe to the promise-pipe, and calling all the methods with wait_=FALSE:

b$DOM$getDocument(wait_ = FALSE) %...>%
  { b$DOM$querySelector(.$root$nodeId, ".sidebar", wait_ = FALSE) } %...>%
  { b$DOM$getBoxModel(.$nodeId, wait_ = FALSE) } %...>%
  str()

# Or, more verbosely:
b$DOM$getDocument(wait_ = FALSE)$
  then(function(value) {
    b$DOM$querySelector(value$root$nodeId, ".sidebar", wait_ = FALSE)
  })$
  then(function(value) {
    b$DOM$getBoxModel(value$nodeId, wait_ = FALSE)
  })$
  then(function(value) {
    str(value)
  })

Each step in the promise chain uses the value from the previous step, via . or value. Note that not all asynchronous code works in such a linear, straightforward way. Sometimes it is necessary to save data from intermediate steps in a broader-scoped variable, if it is to be used in a later step in the promise chain.

Turning asynchronous code into synchronous code

There may be times, especially when programming with Chromote, where you want to wait for a promise to resolve before continuing. To do this, you can use the Chromote or ChromoteSession’s wait_for() method.

# A promise chain
p <- b$DOM$getDocument(wait_ = FALSE) %...>%
  { b$DOM$querySelector(.$root$nodeId, ".sidebar", wait_ = FALSE) } %...>%
  { b$DOM$getBoxModel(.$nodeId, wait_ = FALSE) } %...>%
  str()

b$wait_for(p)
#> List of 1
#>  $ model:List of 6
#>   ..$ content:List of 8
#>   .. ..$ : num 128
#>   .. ..$ : int 28
#>   .. ..$ : num 292
#>   .. ..$ : int 28
#>   .. ..$ : num 292
#>   .. ..$ : num 988
#>   .. ..$ : num 128
#>   .. ..$ : num 988
#>   ..$ padding:List of 8
#>   .. ..$ : num 112
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : num 988
#>   .. ..$ : num 112
#>   .. ..$ : num 988
#>   ..$ border :List of 8
#>   .. ..$ : num 112
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : num 988
#>   .. ..$ : num 112
#>   .. ..$ : num 988
#>   ..$ margin :List of 8
#>   .. ..$ : int 15
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : num 1030
#>   .. ..$ : int 15
#>   .. ..$ : num 1030
#>   ..$ width  : int 195
#>   ..$ height : int 960

This documentation will refer to this technique as synchronizing asynchronous code. The way that wait_for() works is that it runs the Chromote object’s private event loop until the promise has resolved. Because the event loop is private, running it will not interfere with the global event loop, which, for example, may used by Shiny to serve a web application.

The $wait_for() method will return the value from the promise, so instead of putting the str() in the chain, you call str() on the value returned by $wait_for():

p <- b$DOM$getDocument(wait_ = FALSE) %...>%
  { b$DOM$querySelector(.$root$nodeId, ".sidebar", wait_ = FALSE) } %...>%
  { b$DOM$getBoxModel(.$nodeId, wait_ = FALSE) }

x <- b$wait_for(p)
str(x)
#> List of 1
#>  $ model:List of 6
#>   ..$ content:List of 8
#>   .. ..$ : num 128
#>   .. ..$ : int 28
#>   .. ..$ : num 292
#>   .. ..$ : int 28
#>   .. ..$ : num 292
#>   .. ..$ : num 988
#>   .. ..$ : num 128
#>   .. ..$ : num 988
#>   ..$ padding:List of 8
#>   .. ..$ : num 112
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : num 988
#>   .. ..$ : num 112
#>   .. ..$ : num 988
#>   ..$ border :List of 8
#>   .. ..$ : num 112
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : num 988
#>   .. ..$ : num 112
#>   .. ..$ : num 988
#>   ..$ margin :List of 8
#>   .. ..$ : int 15
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : int 28
#>   .. ..$ : num 308
#>   .. ..$ : num 1030
#>   .. ..$ : int 15
#>   .. ..$ : num 1030
#>   ..$ width  : int 195
#>   ..$ height : int 960

There are some methods in Chromote and ChromoteSession objects which are written using asynchronous method calls, but conditionally use wait_for() so that they can be called either synchronously or asynchronously. The $screenshot() method works this way, for example. You can call b$screenshot(wait_=TRUE) (which is the default) for synchronous behavior, or b$screenshot(wait_=FALSE) for async behavior.

If you want to write a function that can be called in either sync or async mode, you can use this basic structure: First, construct a promise chain by calling the CDP methods with wait_=FALSE. Then, at the end, if the user used wait_=TRUE, wait for the promise to resolve; otherwise, simply return the promise.

getBoxModel <- function(b, selector = "html", wait_ = TRUE) {
  p <- b$DOM$getDocument(wait_ = FALSE) %...>%
    { b$DOM$querySelector(.$root$nodeId, selector, wait_ = FALSE) } %...>%
    { b$DOM$getBoxModel(.$nodeId, wait_ = FALSE) }

  if (wait_) {
    b$wait_for(p)
  } else {
    p
  }
}

# Synchronous call
str(getBoxModel(b, ".sidebar"))

# Asynchronous call
getBoxModel(b, ".sidebar", wait_ = FALSE) %...>%
  str()

But, you might be wondering, if we want a synchronous API, why not simply write the synchronous code by calling the individual methods synchronously, and using a normal pipe to connect them, as in:

b$DOM$getDocument() %>%
  { b$DOM$querySelector(.$root$nodeId, ".sidebar") } %>%
  { b$DOM$getBoxModel(.$nodeId) } %>%
  str()

There are two reasons for this. The first is that this would require a duplication of all the code for the sync and async code paths. Another reason is that the internal async code can be written to send multiple independent command chains to the ChromoteSession (or multiple ChromoteSessions), and they will be executed concurrently. If there are multiple promise chains, you can do something like the following to wait for all of them to resolve:

# Starting with promises p1, p2, and p3, create a promise that resolves only
# after they have all been resolved.
p <- promise_all(p1, p2, p3)
b$wait_for(p)

Async events

In addition to commands The Chrome DevTools Protocol also has events. These are messages that are sent from the browser to the R process when various browser events happen.

As an example, it can be a bit tricky to find out when to take a screenshot. When you send the browser a command to navigate to a page, it sends a response immediately, but it may take several more seconds for it to actually finish loading that page. When it does, the Page.loadEventFired event will be fired.

b <- ChromoteSession$new()

# Navigate and wait for Page.loadEventFired.
# Note: these lines are put in a single code block to ensure that there is no
# idle time in between.
{
  b$Page$navigate("https://www.r-project.org/")
  str(b$Page$loadEventFired())
}
#> List of 1
#>  $ timestamp: num 683

With the synchronous API, the call to b$Page$loadEventFired() will block until Chromote receives a Page.loadEventFired message from the browser. However, with the async promise API, you would write it like this:

b$Page$navigate("https://www.r-project.org/", wait_ = FALSE) %...>%
  { b$Page$loadEventFired(wait_ = FALSE) } %...>%
  { str(.) }

# Or, more verbosely:
b$Page$navigate("https://www.r-project.org/", wait_ = FALSE)$
  then(function(value) {
    b$Page$loadEventFired(wait_ = FALSE)
  })$
  then(function(value) {
    str(value)
  })

There will be a short delay after running the code before the value is printed.

However, even this is not perfectly reliable, because in some cases, the browser will navigate to the page before it receives the loadEventFired command from Chromote. If that happens, the load even will have already happened before the browser starts waiting for it, and it will hang. The correct way to deal with this is to issue the loadEventFired command before navigating to the page, and then wait for the loadEventFired promise to resolve.

# This is the correct way to wait for a page to load with async and then chain more commands
p <- b$Page$loadEventFired(wait_ = FALSE)
b$Page$navigate("https://www.r-project.org/", wait_ = FALSE)

# A promise chain of more commands after the page has loaded
p$then(function(value) {
  str(value)
})

If you want to block until the page has loaded, you can once again use wait_for(). For example:

p <- b$Page$loadEventFired(wait_ = FALSE)
b$Page$navigate("https://www.r-project.org/", wait_ = FALSE)

# wait_for returns the last value in the chain, so we can call str() on it
str(b$wait_for(p))
#> List of 1
#>  $ timestamp: num 683

Technical note: The Chrome DevTools Protocol itself does not automatically enable event notifications. Normally, you would have to call the Page.enable method to turn on event notifications for the Page domain. However, Chromote saves you from needing to do this step by keeping track of how many callbacks there are for each domain. When the number of event callbacks for a domain goes from 0 to 1, Chromote automatically calls $enable() for that domain, and when it goes from 1 to 0, it it calls $disable().

In addition to async events with promises, they can also be used with regular callbacks. For example:

b$Page$loadEventFired(callback_ = str)

This tells Chromote to call str() (which prints to the console) on the message value every single time that a Page.loadEventFired event message is received. It will continue doing this indefinitely. (Calling an event method this way also increments the event callback counter.)

When an event method is called with a callback, the return value is a function which will cancel the callback, so that it will no longer fire. (The canceller function also decrements the event callback counter. If you lose the canceller function, there is no way to decrement the callback counter back to 0.)

cancel_load_event_callback <- b$Page$loadEventFired(callback_ = str)

# Each of these will cause the callback to fire.
n1 <- b$Page$navigate("https://www.r-project.org/")
n2 <- b$Page$navigate("https://cran.r-project.org/")

cancel_load_event_callback()

# No longer causes the callback to fire.
n3 <- b$Page$navigate("https://www.rstudio.com/")

Resource cleanup and garbage collection

When Chromote starts a Chrome process, it calls Chrome$new(). This launches the Chrome process it using processx::process(), and enables a supervisor for the process. This means that if the R process stops, the supervisor will detect this and shut down any Chrome processes that were registered with the supervisor. This prevents the proliferation of Chrome processes that are no longer needed.

The Chromote package will, by default, use a single Chrome process and a single Chromote object, and each time ChromoteSession$new() is called, it will spawn them from the Chromote object. See The Chromote object model for more information.

Specifying which browser to use

Chromote will look in specific places for the Chrome web browser, depending on platform. This is done by the chromote:::find_chrome() function.

If you wish to use a different browser from the default, you can set the CHROMOTE_CHROME environment variable, either with Sys.setenv(CHROMOTE_CHROME="/path/to/browser").

Sys.setenv(CHROMOTE_CHROME = "/Applications/Chromium.app/Contents/MacOS/Chromium")

b <- ChromoteSession$new()
b$view()
b$Page$navigate("https://www.whatismybrowser.com/")

Another way is create a Chromote object and explicitly specify the browser, then spawn ChromoteSessions from it.

m <- Chromote$new(
  browser = Chrome$new(path = "/Applications/Chromium.app/Contents/MacOS/Chromium")
)

# Spawn a ChromoteSession from the Chromote object
b <- m$new_session()
b$Page$navigate("https://www.whatismybrowser.com/")

Yet another way is to create a Chromote object with a specified browser, then set it as the default Chromote object.

m <- Chromote$new(
  browser = Chrome$new(path = "/Applications/Chromium.app/Contents/MacOS/Chromium")
)

# Set this Chromote object as the default. Then any
# ChromoteSession$new() will be spawned from it.
set_default_chromote_object(m)
b <- ChromoteSession$new()
b$view()
b$Page$navigate("https://www.whatismybrowser.com/")

Chrome on remote hosts

Chromote can control a browser running on a remote host. To start the browser, open a terminal on the remote host and run one of the following, depending on your platform:

Warning: Depending on how the remote machine is configured, the Chrome debug server might be accessible to anyone on the Internet. Proceed with caution.

# Mac
"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" --headless \
  --remote-debugging-address=0.0.0.0 --remote-debugging-port=9222

# Linux
google-chrome --headless --remote-debugging-address=0.0.0.0 --remote-debugging-port=9222

# Windows
"C:\Program Files (x86)\Google\Chrome\Application\chrome.exe"  --headless \
  --remote-debugging-address=0.0.0.0 --remote-debugging-port=9222

Then, in your local R session, create a Chromote object with the host and port (you will need to use the correct IP address). Once it’s created, you can spawn a session off of it which you can control as normal:

r <- Chromote$new(
  browser = ChromeRemote$new(host = "10.0.0.5", port = 9222)
)

b <- r$new_session()

b$Browser$getVersion()
b$view()
b$Page$navigate("https://www.whatismybrowser.com/")
b$Page$loadEventFired()
b$screenshot("browser.png")
b$screenshot("browser_string.png", selector = ".string-major")

When you use $view() on the remote browser, your local browser may block scripts for security reasons, which means that you won’t be able to view the remote browser. If your local browser is Chrome, there will be a shield-shaped icon in the location bar that you can click in order to enable loading the scripts. (Note: Some browsers don’t seem to work at all with the viewer.)

Technical note: There seem to be some timing issues with remote browsers. In the example above, the browser may finish navigating to the web site before the R process receives the response message for $navigate(), and therefore before R starts waiting for Page.loadEventFired. In order to avoid these timing problems, it is better to write code like this:

{
  p <- b$Page$loadEventFired(wait_ = FALSE)
  b$Page$navigate("https://www.whatismybrowser.com/", wait_ = FALSE)
  b$wait_for(p)
}
b$screenshot("browser.png")

This tells it to fire off the Page.navigate command and not wait for it, and then immediately start waiting for Page.loadEventFired event.

Attaching to existing tabs

In the examples above, we connected to an existing browser, but created a new tab to attach to. It’s also possible to attach to an existing browser and and existing tab. In Chrome debugging terminology a tab is called a “Target”, and there is a command to retrieve the list of current Targets:

r$Target$getTargets()

Every target has a unique identifier string associated with it called the targetId; "9DAE349A3A533718ED9E17441BA5159B" is an example of one.

Here we define a function that retrieves the ID of the first Target (tab) from a Chromote object:

first_id <- function(r) {
  ts <- r$Target$getTargets()$targetInfos
  stopifnot(length(ts) > 0)
  r$Target$getTargets()$targetInfos[[1]]$targetId
}

The following code shows an alert box in the first tab, whatever it is:

rc <- ChromeRemote$new(host = "localhost", port = 9222)
r <- Chromote$new(browser = rc)
tid <- first_id(r)
b <- r$new_session(targetId = tid)
b$Runtime$evaluate('alert("this is the first tab")')

Examples

Loading a page reliably

In many cases, the commands Page$navigate() and then $Page$loadEventFired() will not reliably block until the page loads. For example:

# Not reliable
b$Page$navigate("https://www.r-project.org/")
b$Page$loadEventFired()  # Block until page has loaded

This is because the browser might successfully navigate to the page before it receives the loadEventFired command from R.

In order to navigate to a page reliably, you must issue the loadEventFired command first in async mode, then issue the navigate command, and then wait for the loadEventFired promise to resolve. (If it has already resolved at this point, then the code will continue.)

# Reliable method 1: for use with synchronous API
p <- b$Page$loadEventFired(wait_ = FALSE)  # Get the promise for the loadEventFired
b$Page$navigate("https://www.r-project.org/", wait_ = FALSE)

# Block until p resolves
b$wait_for(p)

# Add more synchronous commands here
b$screenshot("browser.png")

The above code uses the async API to do the waiting, but then assumes that you want to write subsequent code with the synchronous API.

If you want to go fully async, then instead of calling wait_for(p), you would simply chain more promises from p, using $then().

# Reliable method 2: for use with asynchronous API
p <- b$Page$loadEventFired(wait_ = FALSE)  # Get the promise for the loadEventFired
b$Page$navigate("https://www.r-project.org/", wait_ = FALSE)

# Chain more async commands after the page has loaded
p$then(function(value) {
  b$screenshot("browser.png", wait_ = FALSE)
})

This is explained in more detail in the Async Events section.

Taking a screenshot of a web page

Take a screenshot of the viewport and display it using the showimage package. This uses Chromote’s $screenshot() method, which wraps up many calls to the Chrome DevTools Protocol.

b <- ChromoteSession$new()

# ==== Semi-synchronous version ====
# Run the next two lines together, without any delay in between.
p <- b$Page$loadEventFired(wait_ = FALSE)
b$Page$navigate("https://www.r-project.org/", wait_ = FALSE)
b$wait_for(p)
b$screenshot(show = TRUE)  # Saves to screenshot.png and displays in viewer

# ==== Async version ====
p <- b$Page$loadEventFired(wait_ = FALSE)
b$Page$navigate("https://www.r-project.org/", wait_ = FALSE)
p$then(function(value) {
    b$screenshot(show = TRUE)
  })

It is also possible to use selectors to specify what to screenshot, as well as the region (“content”, “border”, “padding”, or “margin”).

# Using CSS selectors, choosing the region, and using scaling
b$screenshot("s1.png", selector = ".sidebar")
b$screenshot("s2.png", selector = ".sidebar", region = "margin")
b$screenshot("s3.png", selector = ".page", region = "margin", scale = 2)

If a vector is passed to selector, it will take a screenshot with a rectangle that encompasses all the DOM elements picked out by the selectors. Similarly, if a selector picks out multiple DOM elements, all of them will be in the screenshot region.

Setting width and height of the viewport (window)

The default size of a ChromoteSession viewport is 992 by 1323 pixels. You can set the width and height when it is created:

b <- ChromoteSession$new(width = 390, height = 844)

b$Page$navigate("https://www.r-project.org/")
b$screenshot("narrow.png")

With an existing ChromoteSession, you can set the size with b$Emulation$setVisibleSize():

b$Emulation$setVisibleSize(width=1600, height=900)
b$screenshot("wide.png")

You can take a “Retina” (double) resolution screenshot by using b$screenshot(scale=2):

b$screenshot("wide-2x.png", scale = 2)

Taking a screenshot of a web page after interacting with it

Headless Chrome provides a remote debugging UI which you can use to interact with the web page. The ChromoteSession’s $view() method opens a regular browser and navigates to the remote debugging UI.

b <- ChromoteSession$new()

b$view()
b$Page$navigate("https://www.google.com") # Or just type the URL in the navigation bar

At this point, you can interact with the web page by typing in text and clicking on things.

Then take a screenshot:

b$screenshot()

Taking screenshots of web pages in parallel

With async code, it’s possible to navigate to and take screenshots of multiple websites in parallel.

library(promises)
library(chromote)
urls <- c(
  "https://www.r-project.org/",
  "https://github.com/",
  "https://news.ycombinator.com/"
)

screenshot_p <- function(url, filename = NULL) {
  if (is.null(filename)) {
    filename <- gsub("^.*://", "", url)
    filename <- gsub("/", "_", filename)
    filename <- gsub("\\.", "_", filename)
    filename <- sub("_$", "", filename)
    filename <- paste0(filename, ".png")
  }

  b <- ChromoteSession$new()
  p <- b$Page$loadEventFired(wait_ = FALSE)
  b$Page$navigate(url, wait_ = FALSE)
  p$then(function(value) {
      b$screenshot(filename, wait_ = FALSE)
    })$
    then(function(value) {
      message(filename)
    })$
    finally(function() {
      b$close()
    })
}

# Screenshot multiple simultaneously
ps <- lapply(urls, screenshot_p)
pa <- promise_all(.list = ps)$then(function(value) {
  message("Done!")
})

# Block the console until the screenshots finish (optional)
cm <- default_chromote_object()
cm$wait_for(pa)
#> www_r-project_org.png
#> github_com.png
#> news_ycombinator_com.png
#> Done!

Setting custom headers

Currently setting custom headers requires a little extra work because it requires Network.enable be called before using it. In the future we’ll streamline things so that it will happen automatically.

b <- ChromoteSession$new()
# Currently need to manually enable Network domain notifications. Calling
# b$Network$enable() would do it, but calling it directly will bypass the
# callback counting and the notifications could get automatically disabled by a
# different Network event. We'll enable notifications for the Network domain by
# listening for a particular event. We'll also store a callback that will
# decrement the callback counter, so that we can disable notifications ater.
disable_network_notifications <- b$Network$responseReceived(function (msg) NULL)
b$Network$setExtraHTTPHeaders(headers = list(
  foo = "bar",
  header1 = "value1"
))

# Visit a web page that prints out the request headers
b$Page$navigate("http://scooterlabs.com/echo")
b$screenshot(show = TRUE)

# Unset extra headers. Note that `list(a=1)[0]` creates an empty _named_ list;
# an empty unnamed list will cause an error because they're converted to JSON
# differently. A named list becomes "{}", but an unnamed list becomes "[]".
b$Network$setExtraHTTPHeaders(headers = list(a=1)[0])

# Request again
b$Page$navigate("http://scooterlabs.com/echo")
b$screenshot(show = TRUE)

# Disable extra headers entirely, by decrementing Network callback counter,
# which will disable Network notifications.
disable_network_notifications()

Custom User-Agent

Synchronous version:

# ==== Synchronous version ====
b$Network$setUserAgentOverride(userAgent = "My fake browser")

b$Page$navigate("http://scooterlabs.com/echo")
b$screenshot(show = TRUE)

# ==== Async version ====
b$Network$setUserAgentOverride(userAgent = "My fake browser", wait_ = FALSE)
p <- b$Page$loadEventFired(wait_ = FALSE)
b$Page$navigate("http://scooterlabs.com/echo", wait_ = FALSE)
p$then(function(value) {
    b$screenshot(show = TRUE)
  })

Extracting text from a web page

One way to extract text from a page is to tell the browser to run JavaScript code that does it:

# ==== Synchronous version ====
b$Page$navigate("https://www.whatismybrowser.com/")

# Run JavaScript to extract text from the page
x <- b$Runtime$evaluate('document.querySelector(".corset .string-major a").innerText')
x$result$value
#> [1] "Chrome 75 on macOS (Mojave)"

# ==== Async version ====
p <- b$Page$loadEventFired(wait_ = FALSE)
b$Page$navigate("https://www.whatismybrowser.com/", wait_ = FALSE)
p$then(function(value) {
    b$Runtime$evaluate(
      'document.querySelector(".corset .string-major a").innerText'
    )
  })$
  then(function(value) {
    print(value$result$value)
  })

Another way is to use CDP commands to extract content from the DOM. This does not require executing JavaScript in the browser’s context, but it is also not as flexible as JavaScript.

# ==== Synchronous version ====
b$Page$navigate("https://www.whatismybrowser.com/")
x <- b$DOM$getDocument()
x <- b$DOM$querySelector(x$root$nodeId, ".corset .string-major a")
b$DOM$getOuterHTML(x$nodeId)
#> $outerHTML
#> [1] "<a href=\"/detect/what-version-of-chrome-do-i-have\">Chrome 75 on macOS (Mojave)</a>"

# ==== Async version ====
p <- b$Page$loadEventFired(wait_ = FALSE)
b$Page$navigate("https://www.whatismybrowser.com/", wait_ = FALSE)
p$then(function(value) {
    b$DOM$getDocument()
  })$
  then(function(value) {
    b$DOM$querySelector(value$root$nodeId, ".corset .string-major a")
  })$
  then(function(value) {
    b$DOM$getOuterHTML(value$nodeId)
  })$
  then(function(value) {
    print(value)
  })

Websites that require authentication

For websites that require authentication, you can use Chromote to get screenshots by doing the following:

Log in interactively and navigate to the page.
Capture cookies from the page and save them.
In a later R session, load the cookies.
Use the cookies in Chromote and navigate to the page.
Take a screenshot.

There are two ways to capture the cookies.

Method 1: The first method uses the headless browser’s viewer. This can be a bit inconvenient because it requires going through the entire login process, even if you have already logged in with a normal browser.

First navigate to the page:

library(chromote)
b <- ChromoteSession$new()
b$view()
b$Page$navigate("https://beta.rstudioconnect.com/content/123456/")

Next, log in interactively via the viewer. Once that’s done, use Chromote to capture the cookies.

cookies <- b$Network$getCookies()
str(cookies)
saveRDS(cookies, "cookies.rds")

After saving the cookies, you can restart R and navigate to the page, using the cookies.

library(chromote)
b <- ChromoteSession$new()
b$view()
cookies <- readRDS("cookies.rds")
b$Network$setCookies(cookies = cookies$cookies)
# Navigate to the app that requires a login
b$Page$navigate("https://beta.rstudioconnect.com/content/123456/")
b$screenshot()

Method 2: The second method captures the cookies using a normal browser. This is can be more convenient because, if you are already logged in, you don’t need to do it again. This requires a Chromium-based browser, and it requires running DevTools-in-DevTools on that browser.

First, navigate to the page in your browser. Then press CMD-Option-I (Mac) or Ctrl-Shift-I (Windows/Linux). The developer tools panel will open. Make sure to undock the developer tools so that they are in their own window. Then press CMD-Option-I or Ctrl-Shift-I again. A second developer tools window will open. (See this SO answer for detailed instructions.)

In the second developer tools window, run the following:

var cookies = await Main.sendOverProtocol('Network.getCookies', {})
JSON.stringify(cookies)

This will return a JSON string representing the cookies for that page. For example:

[{"cookies":[{"name":"AWSALB","value":"T3dNdcdnMasdf/cNn0j+JHMVkZ3RI8mitnAggd9AlPsaWJdsfoaje/OowIh0qe3dDPiHc0mSafe5jNH+1Aeinfalsd30AejBZDYwE","domain":"beta.rstudioconnect.com","path":"/","expires":1594632233.96943,"size":130,"httpOnly":false,"secure":false,"session":false}]}]

Copy that string to the clipboard. In your R session, you can paste it to this code, surrounded by single-quotes:

cookie_json <- '[{"cookies":[{"name":"AWSALB","value":"T3dNdcdnMasdf/cNn0j+JHMVkZ3RI8mitnAggd9AlPsaWJdsfoaje/OowIh0qe3dDPiHc0mSafe5jNH+1Aeinfalsd30AejBZDYwE","domain":"beta.rstudioconnect.com","path":"/","expires":1594632233.96943,"size":130,"httpOnly":false,"secure":false,"session":false}]}]'

cookies <- jsonlite::fromJSON(cookie_json, simplifyVector = FALSE)[[1]]

Then you can use Chromote to navigate to the page and take a screenshot.

library(chromote)
b <- ChromoteSession$new()
b$view()
b$Network$setCookies(cookies = cookies$cookies)
b$Page$navigate("https://beta.rstudioconnect.com/content/123456/")
b$screenshot()

rstudio / chromote

readme