dave yarwood
Alda and the nREPL protocol


December 14, 2020


As you may have heard, I have been busy, working on Alda 2, a ground-up rewrite of Alda using Go and Kotlin. As part of this project, I have re-thought the Alda REPL experience to make it closer in spirit to the nREPL experience that I’ve been enjoying as a Clojure programmer.

The big idea about nREPL is that it is a REPL that operates over a network connection. After starting an nREPL server, any number of nREPL clients can connect to the server and send over code to be evaluated. The server treats each client connection as a separate context, and sends evaluation results back to the client in response.

As I was designing the next major version of Alda, I came to the conclusion that the Alda REPL should function in the same way. The exciting idea that I have in mind is that multiple Alda composers could connect to the same Alda REPL server and compose music together by interacting with the server in real time.

So, I started thinking about the protocol for these interactions between REPL clients and REPL servers. I had read that nREPL is a language-agnostic protocol (i.e. not specific to Clojure) and that it has been used to successfully implement nREPL servers for other languages. So, I decided to try my hand at using the nREPL protocol as the basis for a new, improved Alda 2 REPL experience.

Shortcomings of Alda 1

The current version of Alda (Alda 1) actually has a client/server architecture already. To start the server, you run alda up, and to start a REPL client session, you run alda repl.

I realized, however, that the client/server idea in Alda 1 extends a little too far into user-land. The end user is forced to worry about whether or not a server is running, even outside of the context of a REPL session when they just want to perform basic evaluations at the command line.

For example, if you don’t have a server running, any attempt to play an Alda score from the command line will fail:

$ alda play -c 'cello: o2 a'
# ... after an awkward pause ...
[27713] ERROR Alda server is down. To start the server, run `alda up`.

I have often wished that Alda had a better out-of-the-box experience. Alda does need to start background processes to play your score in an asynchronous manner, but it would be nice if Alda could start these background processes for me, instead of expecting me to run alda up and wait for the server to come up before I can do anything. It would also be nice if Alda could do more of the work that it does without needing to talk to a server at all. For example, the alda CLI should be able to tell me if my score has a syntax error without needing to talk to a background process.

So, I’ve implemented it that way for Alda 2, and now I feel like the basic, everyday usage of the Alda CLI is much more comfortable and satisfying! When Alda 2 is released, users will no longer need to know about Alda’s background processes at all. After downloading and installing Alda, users can run alda play -c 'some code' or alda play -f some-file.alda whenever the inspiration strikes them and Alda will play their score, even though they never ran alda up (a command that no longer exists in Alda 2). There are still background processes out of necessity (an important goal of Alda is that playback is asynchronous), but now, the processes are entirely managed “behind the scenes” by Alda.

The Alda 2 REPL

Whereas the Alda 1 REPL requires you to start an Alda server first (alda up), the Alda 2 REPL automatically starts a REPL server for you when you run alda repl. The interactive REPL session feels a lot like the command line alda play -c '...' experience, except that the REPL remembers the context of your previous evaluations. For example, if you evaluate the following line of code:

accordion: o5 c d e f

You will hear an accordion playing the notes C, D, E, and F in the fifth octave. Then, if you evaluate the following line of code:


You will hear the note G, still on an accordion, still in the fifth octave.

The Alda 2 REPL loads instantaneously, and you can immediately start entering lines of input into your REPL session after you run alda repl. It doesn’t feel like a client/server setup, but under the hood, it is.

It might not look like much now, but this improved client/server REPL architecture will unlock new and exciting possibilities for Alda composers to collaborate in real time.

Imagine a scenario where one composer is sharing their system’s audio with others. Maybe they are performing live on a stage, and their audio is being projected over a sound system. Maybe it’s a virtual event with an audience. Or maybe it’s just a small group of friends on a Zoom call. In any of these scenarios, the composer who’s sharing their audio can start an Alda REPL server (alda repl --server --port 12345) and the other composers can participate by connecting to the REPL server (alda repl --client --host --port 12345). Now, all of the composers have access to the same audio programming environment, and they can interact with the same score, writing their own contributions and hearing them played together.

Just talking about this is giving me goosebumps! I hope that someday, I can make this dream of live, collaborative Alda programming a reality. But for now, I’d like to talk a little more about the nREPL protocol and the fun that I had implementing it for Alda 2.

The nREPL protocol

Messages and transport

nREPL messages are typically sent bencoded over sockets. Bencode is a simple encoding format developed by BitTorrent with existing library support in a number of languages. I tried several bencode libraries for Go, and I ended up choosing this one because it has a simple, intuitive API and it worked out of the box.

Every nREPL message represents either a request from a client or a response from a server. In either case, the message is an associative data structure (think Clojure maps, Ruby hashes, Go maps, JSON objects, etc.). Every request message contains at least an operation (op), and often other entries whose names vary depending on the operation.

Here’s an example request from a client to evaluate a string of code:

  "op": "eval",
  "code": "(+ 1 2 3)"

Strictly speaking, this is all that’s required in a request. However, there are a couple of other things that it is standard for a client request to include:

  • id: a unique identifier (e.g. a UUID) representing the request. The server parrots the id value back in the response so that the client can be sure that the response it’s supposed to receive wasn’t mixed up with the response for another request.

  • session: another unique identifier that represents the session that the client and server are currently participating in. I’ll talk more about sessions below.

Response messages also include id and session entries, and they also may include a status entry, which is actually a list of strings that describe the status of the request. If a request was handled successfully, the status will be ["done"]. If there was an error, the status will be ["done", "error"]. If a response is missing status information altogether, it means that the request is still being processed, and the client should continue to receive responses from the server until it receives one where the status entry includes "done".

Here’s an example response from a server with an evaluation result:

  "id": "d2fa0626-58a3-4abc-b0af-a8afd8b818ad",
  "session": "bb05e357-d9c0-49d2-b206-067c88913e68",
  "status": ["done"],
  "value": "6"

Client/server interactions

The official nREPL documentation about building servers and building clients is concise and super informative. I was basically able to read and implement all of it over the course of a few days, and I had a lot of fun doing it!

When an nREPL client starts, it first sends a clone request to the server to create a new session:

  "id": "dc0a4fb1-0a30-483c-8384-de166cb9bf4d",
  "op": "clone"

The server sends a response that indicates the session was created successfully and includes the ID of the new session:

  "id": "dc0a4fb1-0a30-483c-8384-de166cb9bf4d",
  "new-session": "b08f4313-7f8a-46a0-8a40-5b8424681dbf",
  "status": ["done"]

Next, the client sends a describe request to verify that the server supports all of the operations it needs to perform:

  "id": "0014c5ec-69bd-4fa1-ad78-1aabde04cc4f",
  "op": "describe",
  "session": "b08f4313-7f8a-46a0-8a40-5b8424681dbf"

The server responds with information about what operations it can perform, as well as what versions of things (e.g. Alda) it is running:

  "id": "0014c5ec-69bd-4fa1-ad78-1aabde04cc4f",
  "ops": {
    "clone": {},
    "describe": {},
    "eval": {},
    "eval-and-play": {},
    "export": {},
    "instruments": {},
    "load": {},
    "new-score": {},
    "replay": {},
    "score-data": {},
    "score-events": {},
    "score-text": {},
    "stop": {}
  "session": "b08f4313-7f8a-46a0-8a40-5b8424681dbf",
  "status": ["done"],
  "versions": {
    "alda": {
      "version-string": "1.99.2"

This is useful because it means the client can fail gracefully if it inadvertently connects to an nREPL server that can’t perform the operations that the client needs. In particular, eval-and-play is an operation that I implemented for the Alda nREPL server, but a different kind of nREPL server (like a Clojure nREPL server) would not support that operation, so the Alda client can easily detect that situation and let the user know.

Trying it out

The most fun part was that I was able to test my implementations of the Alda nREPL client and server by pointing them at a Clojure nREPL server and client. This taught me a lot about what an nREPL client or server written “to spec” should do.

The cool thing is that I was able to make my Alda nREPL server functional enough to participate in a Clojure nREPL client session. The Alda nREPL server doesn’t really support eval (at least, not yet?), so it just responds with a shrug emoji, but it’s still fun to see the communication happening:


I had read that the nREPL protocol was simple enough to be easily implemented from scratch, and I really did find that to be the case. I found implementing an nREPL client and server in Go to be a fun little project. The nREPL documentation is very good and easy to follow.

If you are a Clojure programmer, I would encourage you to try the exercise of implementing an nREPL client or server from scratch. In the process, you will discover how the nREPL protocol works, you’ll be surprised to learn just how simple it is, and maybe you’ll build something interesting!


Reply to this tweet with any comments, questions, etc.!