BSD-3-Clause licensed by Dustin Sallings
Maintained by [email protected]
This version can be pinned in stack with:net-mqtt-,4107

Module documentation for


An MQTT protocol implementation for Haskell.

Client Examples


import Network.MQTT.Client
import Network.URI (parseURI)

main :: IO ()
main = do
  let (Just uri) = parseURI "mqtt://"
  mc <- connectURI mqttConfig uri
  publish mc "tmp/topic" "hello!" False


import Network.MQTT.Client
import Network.URI (parseURI)

main :: IO ()
main = do
  let (Just uri) = parseURI "mqtt://"
  mc <- connectURI mqttConfig{_msgCB=SimpleCallback msgReceived} uri
  print =<< subscribe mc [("tmp/topic1", subOptions), ("tmp/topic2", subOptions)] []
  waitForClient mc   -- wait for the the client to disconnect

    msgReceived _ t m p = print (t,m,p)


Changelog for net-mqtt

Ping time is configurable.

Minor build changes and code cleanup. Includes LTS update.

Quick release for updating LTS and deps. Wasn’t committed and pushed because I got lazy.

Fix for pub flow messages with 0 length properties that are explicitly encoded.

Thanks to Marc Jakobi for pointing this out.

A fix allowing the filter to wildcards to match their parent as per spec.

This is not very intuitive and kind of seems like a bad idea, but the spec says this is how it works.

Thanks to Noah Halford.

Added OrderedCallback and OrderedLowLevelCallback to guarantee processing the messages in order.

SimpleCallback and LowLevelCallback both run callbacks in isolated threads such that when a message takes longer to process than another, a second message might finish before the first. Most of the time, this probably doesn’t matter this model is simpler and keeps latency lower. Now applications can decide what’s appropriate for them.

Added toFilter to convert Topics to Filters.

(Thanks Matt Friede)

Exported arbitraryFilter for more control over filter creation.

Added an Arbitrary instance for Filter.

The Topic and Filter types are now newtype wrappers around Text instead of being aliases for text. This is a breaking change, but makes a lot of bugs harder to express. There’s a split function available that will split a Topic or Filter into components, and both types are Semigroups joining on /.

Because these are newtypes, we the Arbitrary instances don’t need special wrappers, so ATopic is gone in favor of just Topic’s Arbitrary instance.

More Arbitrary topic helpers.

Commandline tool enhancements.

Bump version bounds.

Fixed an error where there’d be an ugly crash in a situation where connections were failing regularly and we detected the failure before a connection thread spun up. I was using undefined for the default thread value because it was intended to be immediately set, but did find a way to get there in a failure storm. It’s a Maybe now.

fail on unexpected packets. I had a print in there from very early on. Proper sequences are covered, but if a broker sends the client an unexpected packet, it’d be good to not just ignore it.

When publishing, “no matching subscribers” should not be considered a failure. It’s also not bubbled up to the caller, but it is returned as an error from the broker to basically say the publish was successful, but nobody cares about the thing you published.

mqtt-watch will automatically reestablish sessions by default without reissuing subscriptions (including auto-generated client IDs).

ConnACKFlags now has a SessionReuse type which makes it very clear whether the session is resuming. I was affected by the boolean blindness of the previous variant myself several times.

Connection _properties is now called _connProperties. I developed a separate net-mqtt-lens package that provides a lens into all properties of all types that have properties, but a bit more consistency here is good.

QoS now has an Ord instance.

A new PktID type alias makes it clear which Word16 values were meant to represent a packet ID.

There are fewer threads in publish handling in both directions. This mostly just simplified things, but it also helped prevent a few races when a lot of values arrived at the same time.

Remove a use of fail that prevents us from building under ghc 8.8.

Support query parameters in ws and wss URLs.

Added function to get original connection ACK response packet.

  • connACK (in IO) and connACKSTM

Added low-level support for correlated responses.

I’m not completely sure how generally useful this interface is, but I’ve been using it in a client that’s implementing an RPC type interface across MQTT for a bit now.

Add connection checks to publish phases.

Having a broker/connection die in the middle of a publish in QoS > 0 could result in a thread waiting indefinitely for the response that would not ever arrive.

Users can now specify TLSSettings for mqtts:// and wss:// connections.

Small bit of refactoring of main threads used by the client. It’s a bit easier to reason about their lifecycle now.

All (at least most) of the threads in use by the client are named so when you’re looking at an eventlog, you can see what’s coming and going.

Export MQTTException (thrown from various internal bits).

Added isConnectedSTM for verifying connection state inside of STM transactions (e.g., verifying you’re connected while also waiting for a value in a TChan).

Also, mqtt-watch reconnects on error now.

Relaxed QuickCheck constraint slightly.

Added connection timeout.

Many changes went into this release.

New Features

  • WebSocket support (ws:// and wss://)
  • Added the mqtt-watch CLI tool (which I use a lot)
  • Lots of work on correctness WRT connection and callback failures.
  • Low-Level callbacks (providing all the details of the published message)
  • Added runMQTTConduit to allow running the client over any Conduit provider.


As part of adding features and improving correctness, I’ve made a few API changes.

  • waitForClient now throws an exception on failure. This greatly simplified usage and has made a variety of my applications more reliable when networks and computers fail.
  • Callbacks are now MessageCallback. This primarily allowed me to separate SimpleCallback (the thing most apps want) and LowLevelCallback (a thing I needed when building an MQTT bridge).
  • Removed runClient and runClientTLS from the API. They don’t provide any value over connectURI or runMQTTConduit.
  • All callbacks are now asynchronous. Before, QoS2 would be by necessity, but a bad callback could cause problems with the machinery, so they’re all independent now. This may not be noticeable in most applications, but it’s something to consider.

The QuickCheck Arbitrary instances are exported in a module now, allowing programs to generate useful tests other implementations of mqtt. I’ve been using this package to test a C implementation.

Also, don’t allow 3.1.1 to generate a password without a username. That’s kind of a weird limitation in the older protocol I’m not sure anyone’s run into, but the spec says not to encode things on the wire that way, so it’s useful for interop testing.

With a few attempts to misuse the library, I found some places where error messages weren’t useful enough. There were still two cases where failures turned into indefinite STM errors instead of more informative errors. 1) when the broker declined your connection and 2) when the broker refused your connection. These are errors at differnet layers, so were addressed differently, but should be informative in both cases now.

Unsubscribe was apparently broken in MQTT 5 as well. I’d never tried to use it, and just happened to notice it wasn’t quite right.

Now with no known issues. was released without consulting the github issues page. There were two open bugs – one had already been fixed in the development of 0.5, but another was still present. named the default subscription options defaultSubOptions, but that’s inconsistent with other defaults, so it was renamed to subOptions. This is technically an API incompatibility being introduced and I wouldn’t normally do that, but the API’s been out for a few hours, so I’m preeptively asking for forgiveness.

Major release for MQTT version 5.

The API is mostly the same, but a list of Property values is passed in and returned from a few different fields.

Subscribe responses are now more detailed in the error case, and also return a [Property].

Connections default to Protocol311 (3.1.1), and all behavior should be backwards compatible in these cases. i.e., you can write code as if it were destined for a v5 broker, but properties won’t be sent and responses will be inferred. If you specify your _protocol as Protocol50 in your MQTTConfig, the new features should all work.

Various bugs were fixed along the path of making v5 compatibility, but I’m pretty sure there’s one left somewhere.

Don’t set a message ID of 0.

This had been working fine for a while, but starting in mosquitto 1.6, the server would just hang up on a subscribe request with a message ID of zero.

Link QoS2 completion thread on subscriber.

An exception from a subscriber callback could be silently dropped without completing the handshake which would then cause the MQTT broker to just stop sending messages to the subscriber. Unfortunately, the broker (at least mosquitto) still responds to pings and doesn’t give any useful notification that it’s no longer sending messages.

Introduced Filter type alias to distinguish from Topic.

Reintroduced timeout management at the protocol layer, dropping a connection when a pong response hasn’t come in in a while (~3x longer than the current 30s ping rate). This was mostly after noticing mosquitto do really weird things where it seemed to just forget about all my active connections (other clients figured that out and dropped and reconnected).

Fixed up github links.

Added Network.MQTT.Topic with match to test Topics against wildcards.

Added connectURI to make it easier to connect to mqtt or mqtts via URI.

No externally visible changes, but a few bug fixes I found when writing an application that published in QoS < 2. QoS 0 would likely not transmit (which is probably fine according to the spec, but not very desirable) and QoS1 didn’t check its ACKs, so it would continue to retry after the server ACKd the message.

API Change

Subscriber callbacks now include the MQTT client as the first argument. This breaks a circular dependency that prevented callbacks from being able to publish messages easily.


Updated to stackage LTS 13.2