diff --git a/DESCRIPTION b/DESCRIPTION
index 3c3ef47..77390b4 100644
--- a/DESCRIPTION
+++ b/DESCRIPTION
@@ -14,12 +14,13 @@ License: MIT + file LICENSE
 URL: https://jsonedit.r-lib.org
 Imports:
     cli,
-    jsonlite,
     rlang (>= 1.1.0),
-    V8
+    V8 (>= 6.0.5.9000)
 Suggests:
     testthat (>= 3.0.0)
 Config/testthat/edition: 3
 Encoding: UTF-8
 Roxygen: list(markdown = TRUE)
 RoxygenNote: 7.3.2
+Remotes:
+    jeroen/V8
diff --git a/R/parse.R b/R/parse.R
index 31427a5..7f3d181 100644
--- a/R/parse.R
+++ b/R/parse.R
@@ -78,8 +78,15 @@ text_parse <- function(text, ..., parse_options = NULL) {
   check_string(text)
   parse_options <- parse_options %||% parse_options()
   check_no_text_parse_errors("parse", text, parse_options = parse_options)
-  out <- js_call(jsonc, "ffi_text_parse", text, parse_options)
-  out
+
+  # In particular, JSON arrays can have mixed types like `[1, "a"]` and we
+  # don't want those to be forcibly simplified to `c("1", "a")` on the way in.
+  #
+  # We don't need this for every function, but when we parse a JSON file
+  # we want predictable output, and for us that means no simplification
+  simplify <- FALSE
+
+  jsonc$call("ffi_text_parse", text, parse_options, simplify = simplify)
 }
 
 #' @rdname parse
@@ -98,8 +105,13 @@ text_parse_at_path <- function(text, path, ..., parse_options = NULL) {
   path <- check_and_normalize_path(path)
   parse_options <- parse_options %||% parse_options()
   check_no_text_parse_errors("parse", text, parse_options = parse_options)
-  out <- js_call(jsonc, "ffi_text_parse_at_path", text, path, parse_options)
-  out
+  jsonc$call(
+    "ffi_text_parse_at_path",
+    text,
+    path,
+    parse_options,
+    simplify = FALSE
+  )
 }
 
 #' @rdname parse
@@ -141,7 +153,7 @@ text_parse_errors <- function(text, ..., parse_options = NULL) {
   check_dots_empty0(...)
   check_string(text, .internal = TRUE)
   parse_options <- parse_options %||% parse_options()
-  js_call(jsonc, "ffi_text_parse_errors", text, parse_options)
+  jsonc$call("ffi_text_parse_errors", text, parse_options, simplify = FALSE)
 }
 
 check_no_text_parse_errors <- function(
@@ -253,46 +265,3 @@ error_code_to_error_message <- function(code) {
 
   lookup[[code]]
 }
-
-# Call a JavaScript function
-#
-# This is `jsonc$call()` but with two changes:
-# - A `NULL` returns visibly
-# - The conversion from JSON does NO simplification
-#
-# In particular, JSON arrays can have mixed types like `[1, "a"]` and we
-# don't want those to be forcibly simplified to `c("1", "a")` on the way in.
-#
-# We don't need this for every function, but when we parse a JSON file
-# we want predictable output, and for us that means no simplification
-js_call <- function(context, fun, ...) {
-  args <- list(...)
-
-  if (!is.null(names(args))) {
-    stop("Named arguments are not supported in JavaScript.")
-  }
-
-  args <- vapply(
-    args,
-    function(x) jsonlite::toJSON(x, auto_unbox = TRUE),
-    character(1)
-  )
-
-  args <- paste(args, collapse = ",")
-  src <- paste0("(", fun, ")(", args, ");")
-
-  out <- context$eval(src, serialize = TRUE, await = FALSE)
-
-  if (is.null(out)) {
-    # A JavaScript `undefined` becomes `NULL` and `fromJSON()` doesn't want that
-    NULL
-  } else {
-    # Otherwise, assume JSON but DON'T SIMPLIFY for stability and predictability
-    jsonlite::fromJSON(
-      out,
-      simplifyVector = FALSE,
-      simplifyDataFrame = FALSE,
-      simplifyMatrix = FALSE
-    )
-  }
-}
diff --git a/inst/js/parse.js b/inst/js/parse.js
index ad6f1e1..5d942e3 100644
--- a/inst/js/parse.js
+++ b/inst/js/parse.js
@@ -29,7 +29,7 @@ function ffi_text_parse_at_path(text, path, parse_options) {
   if (node) {
     return get_node_value(node);
   } else {
-    return undefined;
+    return null;
   }
 }
 
diff --git a/tests/testthat/test-parse.R b/tests/testthat/test-parse.R
index 87c9e4c..b69d032 100644
--- a/tests/testthat/test-parse.R
+++ b/tests/testthat/test-parse.R
@@ -3,6 +3,25 @@ test_that("empty file parsing works", {
   expect_identical(text_parse_at_path("", "a"), NULL)
 })
 
+test_that("Output is always returned visibly", {
+  expect_identical(withVisible(text_parse("{}"))$visible, TRUE)
+  expect_identical(
+    withVisible(text_parse_at_path('{ "a": 1 }', "a"))$visible,
+    TRUE
+  )
+
+  # These must return visible `NULL`
+  expect_identical(withVisible(text_parse(""))$visible, TRUE)
+  expect_identical(withVisible(text_parse_at_path("", "a"))$visible, TRUE)
+
+  # These must return visible `NULL`
+  expect_identical(withVisible(text_parse("null"))$visible, TRUE)
+  expect_identical(
+    withVisible(text_parse_at_path('{ "a": null }', "a"))$visible,
+    TRUE
+  )
+})
+
 test_that("works outside of a base object `{`", {
   expect_identical(text_parse("1"), 1L)
   expect_identical(text_parse("1.5"), 1.5)