Keep in mind that the plural form of "middleware" is also "middleware".

Middleware are higher-order functions that accept a handler and return a new handler that may compose additional functionality onto or around the original. For example, some middleware that handles a hypothetical "time?" :op by replying with the local time on the server:

 '[nrepl.misc :refer (response-for)]
 '[nrepl.transport :as t])

(defn current-time
  (fn [{:keys [op transport] :as msg}]
    (if (= "time?" op)
      (t/send transport (response-for msg :status :done :time (System/currentTimeMillis)))
      (h msg))))

A little silly, but this pattern should be familiar to you if you have implemented Ring middleware before. Nearly all of the same patterns and expectations associated with Ring middleware should be applicable to nREPL middleware.

All of nREPL’s provided default functionality is implemented in terms of middleware, even foundational bits like session and eval support. This default middleware "stack" aims to match and exceed the functionality offered by the standard Clojure REPL, and is available at nrepl.server/default-middleware. Concretely, it consists of a number of middleware functions' vars that are implicitly merged with any user-specified middleware provided to nrepl.server/default-handler. To understand how that implicit merge works, we’ll first need to talk about middleware "descriptors".

(See this documentation listing for details as to the operations implemented by nREPL’s default middleware stack, what each operation expects in request messages, and what they emit for responses.)

Middleware descriptors and nREPL server configuration

It is generally the case that most users of nREPL will expect some minimal REPL functionality to always be available: evaluation (and the ability to interrupt evaluations), sessions, file loading, and so on. However, as with all middleware, the order in which nREPL middleware is applied to a base handler is significant; e.g., the session middleware’s handler must look up a user’s session and add it to the message map before delegating to the handler it wraps (so that e.g. evaluation middleware can use that session data to stand up the user’s dynamic evaluation context). If middleware were "just" functions, then any customization of an nREPL middleware stack would need to explicitly repeat all of the defaults, except for the edge cases where middleware is to be appended or prepended to the default stack.

To eliminate this tedium, the vars holding nREPL middleware functions may have a descriptor applied to them to specify certain constraints in how that middleware is applied. For example, the descriptor for the nrepl.middleware.session/add-stdin middleware is set thusly:

(set-descriptor! #'add-stdin
  {:requires #{#'session}
   :expects #{"eval"}
   :handles {"stdin"
             {:doc "Add content from the value of \"stdin\" to *in* in the current session."
              :requires {"stdin" "Content to add to *in*."}
              :optional {}
              :returns {"status" "A status of \"need-input\" will be sent if a session's *in* requires content in order to satisfy an attempted read operation."}}}})

Middleware descriptors are implemented as a map in var metadata under a :nrepl.middleware/descriptor key. Each descriptor can contain any of three entries:

  • :requires, a set containing strings or vars identifying other middleware that must be applied at a higher level than the middleware being described. Var references indicate an implementation detail dependency; string values indicate a dependency on any middleware that handles the specified :op.

  • :expects, the same as :requires, except the referenced middleware must exist in the final stack at a lower level than the middleware being described.

Another way to think of :expects and :requires would be before and after. Middleware you’re expecting should have already been applied by the time the middleware that expects it gets applied, and middleware that’s required should be applied afterwards. We’ll expand on this in the paragraphs to come.
  • :handles, a map that documents the operations implemented by the middleware. Each entry in this map must have as its key the string value of the handled :op and a value that contains any of four entries:

    • :doc, a human-readable docstring for the middleware

    • :requires, a map of slots that the handled operation must find in request messages with the indicated :op

    • :optional, a map of slots that the handled operation may utilize from the request messages with the indicated :op

    • :returns, a map of slots that may be found in messages sent in response to handling the indicated :op

The values in the :handles map are used to support the "describe" operation, which provides "a machine- and human-readable directory and documentation for the operations supported by an nREPL endpoint" (see nrepl.impl.docs/generate-ops-info and the results of lein with-profile +maint run -m nrepl.impl.docs here).

There’s also lein with-profile +maint run -m nrepl.impl.docs --output md if you’d like to generate an ops listing in Markdown format.

The :requires and :expects entries control the order in which middleware is applied to a base handler. In the add-stdin example above, that middleware will be applied after any middleware that handles the "eval" operation, but before the nrepl.middleware.session/session middleware. In the case of add-stdin, this ensures that incoming messages hit the session middleware (thus ensuring that the user’s dynamic scope — including in — has been added to the message) before the add-stdin's handler sees them, so that it may append the provided stdin content to the buffer underlying in. Additionally, add-stdin must be "above" any eval middleware, as it takes responsibility for calling clojure.main/skip-if-eol on in prior to each evaluation (in order to ensure functional parity with Clojure’s default stream-based REPL implementation).

The specific contents of a middleware’s descriptor depends entirely on its objectives: which operations it is to implement/define, how it is to modify incoming request messages, and which higher- and lower-level middleware are to aid in accomplishing its aims.

nREPL uses the dependency information in descriptors in order to produce a linearization of a set of middleware; this linearization is exposed by nrepl.middleware/linearize-middleware-stack, which is implicitly used by nrepl.server/default-handler to combine the default stack of middleware with any additional provided middleware vars. The primary contribution of default-handler is to use nrepl.server/unknown-op as the base handler; this ensures that unhandled messages will always produce a response message with an :unknown-op :status. Any handlers otherwise created (e.g. via direct usage of linearize-middleware-stack to obtain a ordered sequence of middleware vars) should do the same, or use a similar alternative base handler.


Each nREPL message is evaluated within a session. There are two types of sessions: ephemeral sessions and long-lived sessions (or registered sessions).

Ephemeral sessions are used for once off processing of a single message. If a message arrives without a session id, one is created and assigned to it. This is discarded after processing the message. There’s no serialisation guarantee with processing of messages in ephemeral sessions (though they are serialized in the current implementation since they run on the server IO thread). However evals run on a dedicated thread so a running eval can’t block another op.

Long-lived sessions provide two things, persistence of values between messages, and a guarantee for serial execution of messages. The only way to create a long-lived session is to clone an existing session (even an ephemeral one).

Sessions persist dynamic vars (collected by get-thread-bindings) against a unique lookup. This allows you to have a different value for *e from different REPL clients (e.g. two separate REPL-y instances). An existing session can be cloned to create a new one, which then can be modified. This allows for copying of existing preferences into new environments.

Sessions become even more useful when different nREPL extensions start taking advantage of them. debug-repl uses sessions to store information about the current breakpoint, allowing debugging of two things separately. piggieback uses sessions to allow host a ClojureScript REPL alongside an existing Clojure one.

An easy mistake is to confuse a session with an id. The difference between a session and id, is that an id is for tracking a single message, and sessions are for tracking remote state. They’re fundamental to allowing simultaneous activities in the same nREPL. For instance - if you want to evaluate two expressions simultaneously you’ll have to do this in separate session, as all requests within the same session are serialized.

Pretty Printing

Pretty printing support was added in nREPL 0.5 and the API is still considered experimental.

nREPL includes a print middleware to print the results of evaluated forms as strings for returning to the client. This enables using libraries like puget to pretty print the evaluation results automatically. The middleware options may be provided in either requests or responses (the former taking precedence over the latter if any options are specified in both). The following options are supported:

  • :nrepl.middleware.print/print: a fully-qualified symbol naming a var whose function to use for printing. Defaults to the equivalent of clojure.core/pr.

    • The var must point to a function of three arguments:

      • value: the value to print.

      • writer: the java.io.Writer to print on.

      • options: a (possibly nil) map of options.

    • Note well that the printing function is expected to not interact with *out* or *err* at all, even rebinding them (e.g. via with-out-str). Output may be printed to either of those streams during its operation – consider the following example:

    (->> [1 2 3]
         (map (fn [n]
                (println n)
    • The result of the expression is (1 2 3), and evaluating it will result in each of the three numbers being printed to *out*. However, because map is lazy, the calls to println will be interleaved with the operation of the printer function. Hence if the printer function is coupled to *out*, its output might be interleaved with that of the calls to println.

      • Technically, map is not fully lazy – it returns a chunked sequence – but the principle still applies.

    • Further, note that clojure.pprint/pprint rebinds *out* internally (even when using its explicit writer arity). It is not possible to prevent the interleaving of output when using clojure.pprint.

  • :nrepl.middleware.print/options: a map of options to pass to the printing function. Defaults to nil.

  • :nrepl.middleware.print/stream?: if logical true, the result of printing each value will be streamed to the client over one or more messages. Defaults to false.

  • :nrepl.middleware.print/buffer-size: size of the buffer to use when streaming results. Defaults to 1024.

    • Note that this only represents an upper bound on the number of bytes per message – the printing function may also call flush on writer, which will result in a response being sent immediately.

  • :nrepl.middleware.print/quota: a hard limit on the number of bytes printed for each value.

    • A status of :nrepl.middleware.print/truncated will be returned by the middleware if the quota is exceeded. In streamed mode, this will be conveyed as a discrete response after the final printing result. Otherwise, it will be added to the status of the response, and additionally the response will include :nrepl.middleware.print/truncated-keys, indicating which keys in the response were truncated.

  • :nrepl.middleware.print/keys: a seq of the keys in the response whose values should be printed. Defaults to [:value] for eval and load-file responses.

{:op "eval"
 :code "(+ 1 1)"
 :nrepl.middleware.print/print 'my.custom/print-value
 :nrepl.middleware.print/options {:print-width 120}
 :nrepl.middleware.print/stream? true
 :nrepl.middleware.print/buffer-size 1024
 :nrepl.middleware.print/quota 8096}

The functionality of the print middleware is reusable by other middleware. If a middleware descriptor’s :requires set contains #'nrepl.middleware.print/wrap-print, then it can expect:

  • Any responses it returns will have its values printed according to the above options, as provided in the request and/or response.

    • For example, to ensure that :value is printed, responses from the eval middleware look like this:

    {:ns "user"
     :value '(1 2 3)
     :nrepl.middleware.print/keys #{:value}}
  • Any requests it handles will contain the key :nrepl.middleware.print/print-fn, whose value is a function that calls the given printer function with the given options – i.e. its signature is [value writer].

Evaluation Errors

nREPL includes a caught middleware which provides a configurable hook for any java.lang.Throwable that should be conveyed interactively (generally by printing to *err*). Like the print middleware, any options may be provided in either requests or responses (the former taking precedence over the latter if any options are specified in both). The following options are supported:

  • :nrepl.middleware.caught/caught: a fully-qualified symbol naming a var whose function to use to convey interactive errors. Must point to a function that takes a java.lang.Throwable as its sole argument. Defaults to clojure.main/repl-caught.

  • :nrepl.middleware.caught/print?: if logical true, the printed value of any interactive errors will be returned in the response (otherwise they will be elided). Delegates to nrepl.middleware.print to perform the printing. Defaults to false.

{:op "eval"
 :code "(/ 1 0)"
 :nrepl.middleware.caught/caught 'my.custom/print-stacktrace
 :nrepl.middleware.caught/print? true}

The functionality of the caught middleware is reusable by other middleware. If a middleware descriptor’s :requires set contains #'nrepl.middleware.caught/wrap-caught, then it can expect:

  • Any returned responses containing the key :nrepl.middleware.caught/throwable will have that key’s corresponding value passed to the hook.

  • Any handled requests will contain the key :nrepl.middleware.caught/caught-fn, whose value is a function that can be called on a java.lang.Throwable to convey errors interactively.


Sideloading support was added in nREPL 0.7 and the API is still considered experimental, and may change.

nREPL includes sideloader middleware. This provides a Java Class Loader that is able to dynamically load classes and resources at runtime by interacting with the nREPL client (as opposed to using the classpath of the JVM hosting nREPL server).

This performs a similar functionality as the load-file operation, where we can load Clojure namespaces (as source files) or Java classes (as bytecode) by simply require or import them.

This allows a client to add new functionality to an already running instance of nREPL, and paves the way for a more client configurable nREPL.

Starting the sideloader

The sideloader is initialised by a new operation: sideloader-start. This will never return with status :done, but its message ID will be used for all future sideloading requests.

As additional resources/classes are looked up, server will send messages to the client with status :sideloader-lookup, and the following parameters

  • :type, being resource or class

  • :name, being a string (e.g. foo/bar.clj or foo.bar)

The client is responsible for responding to the lookup. It does so by replying with the operation sideloader-provide, and parameters

  • :type and :name the same as the lookup.

  • :content base-64 encoded byte array of the resource or class. An empty response indicates "resource/class not found."

See Building Clients on how to implement the server side of this functionality.

Triggering the sideloader

Once the sideloader has been started, it can be triggered by an ordinary eval or load-file operation. The nREPL server will first try to find a resource/class on the classpath of its own JVM. Failing this, it will attempt to request it from the nREPL client by the above described mechanism. Should this still fail, expect a ClassNotFoundException as usual.

Once a class has been loaded, it will become available to all sessions on nREPL server.

Dynamic middleware loading

Dynamic middleware loading support was added in nREPL 0.8 and the API is still considered experimental, and may change.

nREPL includes a dynamic-loader middleware, which can be used, at runtime, query and change the middleware stack the nREPL server is using. This is especially powerful when combined with sideloading, as it allows a client to configure the server after connecting, and provides an alternative to having to specify the middleware required by the client as startup time.

This introduces three new operations:

  • ls-middleware, to return a list of active middleware, ordered from inside outwards.

  • add-middleware, which adds a middleware to the stack. Optionally, a list of extra-namespaces could be provided for loading. This is useful when adding middleware that implement some form of deferred loading. Examples include cider-nrepl and refactor-nrepl. In these cases, some of the required namespaces might only be loaded upon first use, which may occur outside of a sideloading session, and thus fail. This feature allows us to pre-load namespaces when we add a middleware. If loading of any particular middleware fails, the stack will be unchanged.

  • swap-middleware, similar to add-, but replaces all existing middleware. Note that this may remove the dynamic-loader itself.