# Chorus is Good Software


I’ve been using [Chorus](https://chorus.sh/) for the past 6-7 months. Within the
first couple days of using it, I was telling everyone I talk with about AI stuff
to try it out. Melty Labs, the company behind Chorus, subsequently built
[Conductor](https://conductor.build/). It appears this Conductor now their
primary focus, and as such they’ve decided to
[open source](https://github.com/meltylabs/chorus) Chorus.

Chorus is a macOS LLM client. Its differentiating feature is that it fetches you
responses from many LLMs in parallel within the same chat. I think it’s quite
good software.

![Chorus screenshot](https://benjamincongdon.me/blog/2025/12/13/Chorus-is-Good-Software/chorus_screenshot.png)

**It is _snappy_.** I used to be impressed by the responsiveness of the ChatGPT
app, but after having become a regular Chorus user, I can’t use either the
ChatGPT[^1] or Claude macOS apps without feeling like they’re just _sluggish_ in
comparison.

**It has a small surface area, but executes the UX well.** The key UX innovation
of Chorus is its multi-model chat -- you can get responses from multiple models,
pick the one you want to thread into the conversation, and continue. There are
also many nice affordances in Chorus that don’t exist in other clients. For
example, chats allow for “inline replies” where you can ask a model about a
response without adding that response to the context window. Chorus also handles
pasted content _much_ better than ChatGPT desktop or the Gemini AI Studio.
(Claude handles pasted stuff pretty well.) Chorus just feels like a _tool_ -- as
opposed to a consumer app -- in a way that other clients don’t.

**It puts the user in control more-so than any of the first-party clients.**
Chat branching and “previous messaging” editing worked in Chorus well before
ChatGPT and Claude.[^2] Using a third-party client like Chorus also lets you
jettison the opinionated system prompts that the first-party clients inject, and
also let you switch quickly between system prompts. The first-party system
prompts are good in a consumer app, but Chorus makes for a better “training
wheels off” LLM client.

**The multi-model native chat makes it easier to build an intuitive sense of
models.** Interacting with Claude/Gemini/GPT-N within the same dialog gives you
a much more intuitive sense of the “shape” of each of the models. It’s also
quite interesting to see, say, Claude Sonnet continuing from a Gemini response,
since their inherent writing styles are quite different.

**It allows you to burn tokens.** First-party chat clients are mostly on a flat,
monthly subscription model. The incentives are to provide you with a good
experience so you don’t cancel, while otherwise minimizing the number of tokens
you use. You pay for all your Chorus tokens. So, want to generate a response
from Opus 4.5 and GPT-5.1 and Sonnet 4.5 and Gemini Pro 3 for each conversation
turn? Go for it. Often times, many fanned-out responses to a message will be
more useful than the “best” model’s response. It’s effectively an easy way to do
“LLM councils” or best-of-N.[^3]

**It’s very clearly not trying to get your data.** Chorus originally shipped
with an optional subscription model that was essentially a model proxy. However,
it was always possible to use your own API keys and their
[privacy policy](https://docs.chorus.sh/account/privacy) was refreshingly clear:
“We don’t look at your chats, and we don’t want to.” Now that Chorus is
open-sourced, you can verify that.

We’re still quite early in figuring out what UX patterns work well for LLMs.
Chorus took an interesting idea, executed it well, and as a result has a really
great novel UX that none of the other clients nail. There are obviously many use
cases for LLMs. The spotlight now is on agentic harnesses like Claude Code,
Codex, Cursor, and the like. Chat interfaces are not solved! Assuredly there is
more space to explore the
[“calculator for words”](https://simonwillison.net/2023/Apr/2/calculator-for-words/)
idea space.

I wish Charlie and team the best of luck on Conductor, and thank them for making
Chorus open source. 🎉 I’m looking forward to contributing & hacking on Chorus.

[^1]:
    The macOS app for ChatGPT feels noticeably sluggish now. I think this
    regression started around the GPT-5 release, if memory serves. My mental
    ranking used to be that the native macOS ChatGPT app was ~2x faster than
    Claude. Now, that’s been reversed.

[^2]:
    Though I will say that at this point, the ChatGPT/Claude desktop apps have
    largely caught up.

[^3]:
    Well, where N = “number of models”, not “number of responses”. I think
    Chorus could add support for multiple generations from the same model, but
    that doesn’t exist today.