diff --git a/sphinx_json_schema_spec/__init__.py b/sphinx_json_schema_spec/__init__.py
new file mode 100644
index 0000000..72c6447
--- /dev/null
+++ b/sphinx_json_schema_spec/__init__.py
@@ -0,0 +1,146 @@
+from contextlib import suppress
+from datetime import datetime
+from importlib import metadata
+from urllib.parse import urljoin
+import os
+import urllib.request
+
+from docutils import nodes
+from lxml import html
+import certifi
+
+__version__ = "1.2.0"
+
+BASE_URL = "https://json-schema.org/draft-07/"
+VALIDATION_SPEC = urljoin(BASE_URL, "json-schema-validation.html")
+REF_URL = urljoin(BASE_URL, "json-schema-core.html#rfc.section.8.3")
+SCHEMA_URL = urljoin(BASE_URL, "json-schema-core.html#rfc.section.7")
+
+
+def setup(app):
+    """
+    Install the plugin.
+
+    Arguments:
+
+        app (sphinx.application.Sphinx):
+
+            the Sphinx application context
+    """
+
+    app.add_config_value("cache_path", "_cache", "")
+
+    os.makedirs(app.config.cache_path, exist_ok=True)
+
+    path = os.path.join(app.config.cache_path, "spec.html")
+    spec = fetch_or_load(path)
+    app.add_role("kw", docutils_does_not_allow_using_classes(spec))
+
+    return dict(version=__version__, parallel_read_safe=True)
+
+
+def fetch_or_load(spec_path):
+    """
+    Fetch a new specification or use the cache if it's current.
+
+    Arguments:
+
+        cache_path:
+
+            the path to a cached specification
+    """
+
+    headers = {
+        "User-Agent": "python-jsonschema v{} - documentation build v{}".format(
+            metadata.version("jsonschema"),
+            __version__,
+        ),
+    }
+
+    with suppress(FileNotFoundError):
+        modified = datetime.utcfromtimestamp(os.path.getmtime(spec_path))
+        date = modified.strftime("%a, %d %b %Y %I:%M:%S UTC")
+        headers["If-Modified-Since"] = date
+
+    request = urllib.request.Request(VALIDATION_SPEC, headers=headers)
+    response = urllib.request.urlopen(request, cafile=certifi.where())
+
+    if response.code == 200:
+        with open(spec_path, "w+b") as spec:
+            spec.writelines(response)
+            spec.seek(0)
+            return html.parse(spec)
+
+    with open(spec_path) as spec:
+        return html.parse(spec)
+
+
+def docutils_does_not_allow_using_classes(spec):
+    """
+    Yeah.
+
+    It doesn't allow using a class because it does annoying stuff like
+    try to set attributes on the callable object rather than just
+    keeping a dict.
+    """
+
+    def keyword(name, raw_text, text, lineno, inliner):
+        """
+        Link to the JSON Schema documentation for a keyword.
+
+        Arguments:
+
+            name (str):
+
+                the name of the role in the document
+
+            raw_source (str):
+
+                the raw text (role with argument)
+
+            text (str):
+
+                the argument given to the role
+
+            lineno (int):
+
+                the line number
+
+            inliner (docutils.parsers.rst.states.Inliner):
+
+                the inliner
+
+        Returns:
+
+            tuple:
+
+                a 2-tuple of nodes to insert into the document and an
+                iterable of system messages, both possibly empty
+        """
+
+        if text == "$ref":
+            return [nodes.reference(raw_text, text, refuri=REF_URL)], []
+        elif text == "$schema":
+            return [nodes.reference(raw_text, text, refuri=SCHEMA_URL)], []
+
+        # find the header in the validation spec containing matching text
+        header = spec.xpath("//h1[contains(text(), '{0}')]".format(text))
+
+        if len(header) == 0:
+            inliner.reporter.warning(
+                "Didn't find a target for {0}".format(text),
+            )
+            uri = VALIDATION_SPEC
+        else:
+            if len(header) > 1:
+                inliner.reporter.info(
+                    "Found multiple targets for {0}".format(text),
+                )
+
+            # get the href from link in the header
+            uri = urljoin(VALIDATION_SPEC, header[0].find("a").attrib["href"])
+
+        reference = nodes.reference(raw_text, text, refuri=uri)
+        return [reference], []
+
+    return keyword