Built in support for newline separated messages via stdin/stdout (for json communication)?

[jakemac53]

This package supports out of the box an IsolateChannel, so it makes sense it could also support a stdin/stdout channel.

One of the most common use cases for that, is JSON protocols, which want to send newline separated JSON encoded messages back and forth.

It is a bit cumbersome today to do that, you need something along these lines:

StreamChannel.withCloseGuarantee(io.stdin, io.stdout)
        .transform(StreamChannelTransformer.fromCodec(utf8))
        .transformStream(const LineSplitter())
        .transformSink(
          StreamSinkTransformer.fromHandlers(
            handleData: (data, sink) {
              sink.add('$data\n');
            },
          ),
        ),

There isn't a ton going on here but it is actually somewhat hard to come up with this magical incantation yourself (I ended up reaching out to @natebosch ).

Note that this implementation assumes each message passed to the sink is a full encoded JSON object.

We could expose this from other packages like package:json_rpc_2 but I think its probably a more general thing?

[jakemac53]

Alternatively the json_rpc_2 package could allow some configuration for arbitrary data to be written to the sink after each encoded message, which might be a more general purpose solution for that particular package (for binary message formats this could be any message termination byte(s)).

[jakemac53]

cc @natebosch

[jakemac53]

Yet another idea could be a JsonObjectLineSplitter which is a sink transformer that just counts open/close curly braces (taking into account escapes/strings etc) and injects newlines between objects.

[lrhn]

We could consider making the JSON encoder and decoder support JSONL when used as a stream transformer.

They're converting between Stream<Object?> and Stream<String>/Stream<Uint8List> (when UTF-8 fused). That API can allow adding or reading more than one object from the same input.

[domesticmouse]

I feel this would be useful for implementing Anthropic Model Context Protocol servers and clients.

I’d like to build a MCP client as a Flutter app with Gemini as the driving LL model (using Vertex AI in Firebase) for an advanced Flutter AI codelab.

Let me know if I can help with proving things out.

/cc @RedBrogdon for visibility

[jakemac53]

I feel this would be useful for implementing Anthropic Model Context Protocol servers and clients.

This is the exact use case I ran into haha (see here).

I’d like to build a MCP client as a Flutter app with Gemini as the driving LL model (using Vertex AI in Firebase) for an advanced Flutter AI codelab.

@natebosch might also be looking into a similar thing fwiw