I was really excited to try this but this does NOT work the way I expected. I wanted a simple git worktree manager for my existing, already-checked-out repository. Instead, it requests Github permissions and clones the repo from Github. This is bad, because you need to run all the dependency installs, etc. for every workspace before being able to test anything. In other words it's like Codex or Cursor Background Agents, except worse, because it's on your machine. The reason I don't use Codex or Background Agents is because my project has way too many dependencies and would take way too long (and too complicated) to install everything required to have an isolated running instance, plus there would be DNS conflicts, external API conflicts, among other issues.
What I do want is simple git worktree management for an already-checked-out repo on my machine, no Github permissions or dependency re-installation (copying node_modules, etc.).
I'm the only user at the moment, and I really enjoy the workflow. I run about four claude-codes at once this way. It's a little underbaked but I think this is the way a lot of people are going to go. Seems like the 'par' tool in a sibling comment is a similar approach.
Containers do make things easier, especially since agents can see the log output of any service super easily. To do the same thing outside a container you need to send logs somewhere the agent can see.
I tried Crystal for the first time today and really enjoyed it. It’s a bit rough around the edges but it’s usable and notably it Does One Thing Really Well. All I wanted was a simple way to run a bunch of Claude codes in parallel and isolation and it does exactly that. No more, no less. No opinionation about how I organize my code, no custom config files I have to think about and manage, just simple work isolation workflow with a GUI.
I tried out git worktrees recently and while they do what's on the tin, I really didn't like them as much as I thought I would. I actually like the way cursor does it despite the fact that it is a VM - I wish I could have their same UI/UX but with local worktrees.
The git worktrees feature has first garnered public attention since anthropic mentioned it in its claude code good practices. Until then it was just one of these obscure git features most professionals have never heard of.
Now all these agents and claude code wrappers are trying to make use of git worktrees - but in all honesty - it's overcomplicating most workflows more than being of a real use for agentic coding. The bottleneck was - for most workflows - never the ai not being able to write 10 different poc of my feature simultaneously - but the human factor - needing to carefully review what ai produced and still steering ai in the right direction. git worktrees doesn't help solve any of these problems. it just adds an unnecessary layer on top.
Yes I had a similar experience. the thing that tripped me up with git worktrees, which is maybe obvious in retrospect, is that they don’t include things that are not tracked by git - e.g. .env.development.local
So starting a new worktree requires additional setup and isn’t as simple as just checking out a new branch
But that is also the selling point of worktrees - I have them for multiple features I am working on, I just spin up services that I need for each feature separately, like databases, etc. How else are you going to isolate migrations, configuration changes, etc.
Sorry about that! We’re working on making this more intuitive.
Internally our workflow looks like this:
- we have a script that sets up a repo — copies env variables, runs pnpm i, inits a db, etc
- we have a field in the repo settings called “setup script.” every time you make a new workspace, that script runs
Hopefully will be much improved over the next week or two!
I'm working on a little wrapper that solves this problem. I have similar needs with .env files, and in my case running 'uv sync' to install dependencies. I linked it elsewhere in this thread so I won't repeat the URL (autowt). It's definitely possible to make this workflow effective with some scripting.
Hi, creator here. If you want to copy node_modules instead of reinstalling, you can click on the repo and add a setup script that does the copy. Sorry it’s a bit obscure—we’re working on making this easier!
Would love to discuss more what kind of testing setup you’re looking for, want to shoot me an email at jackson@melty.sh?
"This is bad, because you need to run all the dependency installs", you must be using Python or one of the other locally cached dependency language tookings. Is never understood why people wanto clone everything every time. Maven and its sibling use ~/.m2 so everything gets cached once. I ceinge every time I have to re-download the universe when I inevitably have to delete my venv and redo anything, what a waste of time and bandwidth and space.
Hey, you might be interested in my project Plandex[1] which uses a shadow git repo approach to keep changes separate until you approve them. It solves the same problem as worktrees.
I just tried Claude Squad this morning , the instructions to use and interface was very clunky . There also was no uninstall instructions or scripts , so I had to write one and uninstall it. Lame
How do you suggest simultaneous independent branches be worked on concurrently from a single repository copy, or is it that you would simply prefer the local work tree method of duplication? And, for example, how do you suggest the same node_modules be used if part of the task requires changing, removing, or updating dependencies? Additionally, how much time do you expect a developer to spend implementing this new system your suggesting across all platforms for the benefit of it taking you a bit less time, especially when pulling dependencies from a local caches as I’m sure you’re doing? Or are you making a suggestion that you plan to contribute?
I’m sure the authors would appreciate well-thought alternative suggestions and assistance.
So this is why Claude Code is so slow now. I am all for these but not at the cost of other more casual users. I never had to worry about Sonnet use but even that is not guaranteed. Forget Opus.
Right now the app uses GitHub's OAuth sign in (https://docs.github.com/en/apps/oauth-apps/building-oauth-ap...) which unfortunately doesn't allow for fine-grained permissions (it will only have access to organization code if you explicitly grant it)
. We're switching our sign-in to a GitHub App so we can make the permissions fine-grained.
Yeah, it’s not like this is a saas and you’d need back-end access to the repos. I suspect this is being run with business potential in mind. The OP would do better by making that clear. And if you are selling a self-hosted app, just charge a license fee. People on 100-200/month claude code subscription wouldn’t mind paying 10-30 bucks for this.
I'm in the midst of modernizing a 1994 K&R C 32-bit codebase to 64-bit C23, as a museum piece: the computer algebra system Macaulay that I coauthored, since replaced by Macaulay 2.
The horror!
On macOS my environment is Sublime Text as editor, iTerm as terminal. I open two full screen iTerm windows of four panes each, start Claude Code Opus 4 and type "just task" in each pane. CLAUDE.md tells the session to run this justfile command and follow the directions, which generally include instructions to continue by doing "just task" again. Each session may edit its <file>.c and <file>.h but not shared headers; if it has a problem with this I touch STOP in the task directory, the other sessions eventually are told to wait, and then the trouble child session can edit what it wants.
I'm sure in a few years AI will read and fix my 57 C file code base all at once, but for now this workflow has proved invaluable.
The irony? This is the dumbest parallel task manager imaginable yet it works. And our code base has a similar feel: The storage management is a set of fixed sized stacks that outperformed malloc() by 4x back in the day. Etc. Etc.
I learned this gift from my Dad. He devised the Bayer filter for digital photography back in 1974, it looks like ten minutes work (it wasn't) but we're still using it.
For me, getting Claude Code working properly two days in a row (12 -14 hours per day) is still a challenge.
Let it running in parallel "unwatched" will end up with nothing but a pile of code that will have to be re-written.
It was not that way at the beginning, when I started in May 24th (2 days post Sonnet/Opus 4 announcement), however, last 2 - 3 weeks, they've tweaked they system substantially (I mean I have no other explanation) and its behavior is despicable.
I have see a few on X complaining the same (along with far more about GPT o3 and 4o) but the culture of "opaque transparency" in the AI industry left us with nothing but guesses.
Who does your design? I absolutely love this aesthetic, both this product, and your site (chorus.sh). What is this even called? I went for a similar vibe with tinylogger.com, but def didn't have your skills to pull it off.
Ah thank you! Our designer is Julian Kelly (https://jfk.works). He's amazing, and worked on Messenger at Meta — a background which has been surprisingly helpful since everything we're working on is a spin on chatting.
Hi, appreciate the note! I'm Julian, I do all the design at Melty. I try to design our software to be functional, visually subtle, and chromatically warm. I don't know that the aesthetic has a name, but I've spent a lot of time studying manuscripts from the early middle ages so that might have seeped in - the logo for Conductor is actually a variant of insular majuscule (you can see the ductus in the animation)
This is very useful! I think the interfaces around models being used in an async manner will look very different than the synchronous chat UIs we are today. Claude Code is the first real “agent” that is providing true economic value, and there’s so much low hanging fruit in making the interfaces far better.
I've been looking for a tool like this, that lets Claude operate on multiple repos. I work on a project that has frontend/packages and backend (separate repos instead of a monorepo for good reasons) and I often develop features that cross both repos. With the terminal I can clone both down into a directory and start Claude code there, but all the tools for background/multiplexing are always built around a single repo. Any chance I can get multi-repo tasks supported?
1. Place all your repos inside a parent directory and name them accordingly (frontend, backend, devops, etc.)
2. Each repo has its own claude file.
3. The parent repo has a claude file that links these claude files.
I’ve had claude successfully navigate across these and uses services with back-end, setup terraform infra and then complete the implementation for the front-end.
That works when you just do it once. Where you can run into problems is when you do that twice. One instance might change a file one way, and the other might change it in a conflicting way, then you have two Claudes getting confused about the state of that file and it gets messy. You can solve this by checking out multiple copies of your repo and having only one instance of Claude working per copy. This app seems like it just provides you with a nice UI for doing that.
Yes that's how I do Claude code manually, git branch per terminal window, and very rarely multiple agents on the same branch
It mostly works, except we don't have a clean flow for docker: shared system daemon & repository means need to manually tag & run by branch/project (docker compose -p ...), which is friction for the LLM and even more setup than we want
As a heavy multi session Claude Code user, this may be what finally converts me to cloud IDEs...
- I switch in between planning and execution in the middle of the conversation using Terminal a lot, it would be nice to have it here as well rather than defining how I want Claude to think in advance.
- Entering messages when the agents can result in task lists to save time.
I’m also experimenting different UX with CC, here is Claude Code running in Slack if anybody is interested. https://peerbot.ai
I have been imagining something like this would be perfect for working with Claude Code. I tried a couple other apps but they seemed to change too much without providing enough value for that change. This feels like just a nice clean simple extension of how Claude code already works that solves my most common pain points in the workflow.
> It'd be great to change the default branch used for creating new workspaces.
Yeah you can actually change this now! If you click the repo name you can make changes to the "setup script". If you added `git checkout -b "branch name"` it would run that on every new workspace instance.
At the moment it's mostly Cursor or VS Code, but I was actually thinking of SourceTree. I'd like to look at the pending changes and manage the commits myself, and I could do that if I could add "open -a SourceTree ." as a custom command. I didn't see a place to edit a setup script, is that just on the filesystem?
Cool idea, and I'm definitely not in the target market (I'm a Linux user and also very hesitant to adopt proprietary tools to my important workflows), but something like this could be useful. Right now though it's pretty easy to run each in a different tmux window and check on it, but once I have about 3 or more running at a time it's very easy to forget about one and have it paused waiting for confirmation. A tool like this to add a dashboard of sorts would be nice. Definitely interesting to think about!
I noticed that I found Claude great for squashing bugs/issues, and started using worktrees but saw similar problems as you described, so I enlisted Claude and wrote https://github.com/Someblueman/hydra to solve for this very issue, keeping (or rather trying) it in pure shell for personal portability and gives me a playground for AI driven development
I am smart, capable, and have a lot of programming experience, and can just about manage to stay focused enough to properly review the output of a single Claude agent.
I'm surprised people are running multiple agents, and are able to check their outputs diligently.
Would be cool if I can use this with opencode, Amazon Q or whatever. I reckon the logic would be quite similar. Seen a few of these tools but they're all Claude-centric.
Full read-write access required to all your Github account's repos. Not just code. Settings, deploy keys. The works. Full access to your organisation settings. Not a privacy policy in sight. Zero disclosure of data practices.
You are INSANE to authorize this app on anything other than throwaway code.
It will only have access to organization code if you explicitly grant it.
We're working on switching our sign-in to a GitHub App so we can make the permissions fine-grained.
Totally understand, but you're asking for the keys for the kingdom, without any data disclosure or privacy policy. Even if you switch to fine-grain permissions, the lack of any "and here are our commitments to handling or accessing your data" is (and should be) a show-stopper for anyone trying this out.
In case people aren't reading, here's the c+v of what Conductor gets access to currently:
--
This application will be able to read and write all public and private repository data. This includes the following:
Note: In addition to repository related resources, the repo scope also grants access to manage organization attributes and organization-owned resources including projects, invitations, team memberships and webhooks. This scope also grants the ability to manage projects owned by users.
What I do want is simple git worktree management for an already-checked-out repo on my machine, no Github permissions or dependency re-installation (copying node_modules, etc.).
I'm the only user at the moment, and I really enjoy the workflow. I run about four claude-codes at once this way. It's a little underbaked but I think this is the way a lot of people are going to go. Seems like the 'par' tool in a sibling comment is a similar approach.
Containers do make things easier, especially since agents can see the log output of any service super easily. To do the same thing outside a container you need to send logs somewhere the agent can see.
So starting a new worktree requires additional setup and isn’t as simple as just checking out a new branch
Internally our workflow looks like this: - we have a script that sets up a repo — copies env variables, runs pnpm i, inits a db, etc - we have a field in the repo settings called “setup script.” every time you make a new workspace, that script runs
Hopefully will be much improved over the next week or two!
Would love to discuss more what kind of testing setup you’re looking for, want to shoot me an email at jackson@melty.sh?
1 - https://github.com/plandex-ai/plandex
works like you'd expect
https://github.com/smtg-ai/claude-squad
I’m sure the authors would appreciate well-thought alternative suggestions and assistance.
Right now the app uses GitHub's OAuth sign in (https://docs.github.com/en/apps/oauth-apps/building-oauth-ap...) which unfortunately doesn't allow for fine-grained permissions (it will only have access to organization code if you explicitly grant it) . We're switching our sign-in to a GitHub App so we can make the permissions fine-grained.
The tools like Cursor have privacy mode that ensure that.
Lots of enterprises use Azure, GCP, AWS for the models for these reasons as well.
The horror!
On macOS my environment is Sublime Text as editor, iTerm as terminal. I open two full screen iTerm windows of four panes each, start Claude Code Opus 4 and type "just task" in each pane. CLAUDE.md tells the session to run this justfile command and follow the directions, which generally include instructions to continue by doing "just task" again. Each session may edit its <file>.c and <file>.h but not shared headers; if it has a problem with this I touch STOP in the task directory, the other sessions eventually are told to wait, and then the trouble child session can edit what it wants.
I'm sure in a few years AI will read and fix my 57 C file code base all at once, but for now this workflow has proved invaluable.
The irony? This is the dumbest parallel task manager imaginable yet it works. And our code base has a similar feel: The storage management is a set of fixed sized stacks that outperformed malloc() by 4x back in the day. Etc. Etc.
I learned this gift from my Dad. He devised the Bayer filter for digital photography back in 1974, it looks like ten minutes work (it wasn't) but we're still using it.
Sadly, this is lost with conductor.
I just don’t feel as joyful using it.
Let it running in parallel "unwatched" will end up with nothing but a pile of code that will have to be re-written.
It was not that way at the beginning, when I started in May 24th (2 days post Sonnet/Opus 4 announcement), however, last 2 - 3 weeks, they've tweaked they system substantially (I mean I have no other explanation) and its behavior is despicable.
I have see a few on X complaining the same (along with far more about GPT o3 and 4o) but the culture of "opaque transparency" in the AI industry left us with nothing but guesses.
https://github.com/stravu/crystal
1. Place all your repos inside a parent directory and name them accordingly (frontend, backend, devops, etc.)
2. Each repo has its own claude file.
3. The parent repo has a claude file that links these claude files.
I’ve had claude successfully navigate across these and uses services with back-end, setup terraform infra and then complete the implementation for the front-end.
But I'm an old school vim/emacs keyboard only hacker, so terminal all the way is super efficient for me
It mostly works, except we don't have a clean flow for docker: shared system daemon & repository means need to manually tag & run by branch/project (docker compose -p ...), which is friction for the LLM and even more setup than we want
As a heavy multi session Claude Code user, this may be what finally converts me to cloud IDEs...
But no way to find out if there’s any data sent to your servers etc, unless I’m missing some links?
- I switch in between planning and execution in the middle of the conversation using Terminal a lot, it would be nice to have it here as well rather than defining how I want Claude to think in advance.
- Entering messages when the agents can result in task lists to save time.
I’m also experimenting different UX with CC, here is Claude Code running in Slack if anybody is interested. https://peerbot.ai
> - Entering messages when the agents can result in task lists to save time.
Ah yes, queuing coming soon too! Do you prefer if the queued messages send when Claude is finished or interrupt it?
Anyways, excellent work!
- It'd be great to change the default branch used for creating new workspaces.
- I'd like the ability to add custom tools to the "Open in..." menu.
> It'd be great to change the default branch used for creating new workspaces. Yeah you can actually change this now! If you click the repo name you can make changes to the "setup script". If you added `git checkout -b "branch name"` it would run that on every new workspace instance.
orig_branch=$(git branch --show-current) && git checkout dev && git pull && git branch -D "$orig_branch" && git checkout -b "$orig_branch"
I'm surprised people are running multiple agents, and are able to check their outputs diligently.
You are INSANE to authorize this app on anything other than throwaway code.
@charlieholtz care to comment?
It will only have access to organization code if you explicitly grant it. We're working on switching our sign-in to a GitHub App so we can make the permissions fine-grained.
In case people aren't reading, here's the c+v of what Conductor gets access to currently:
--
This application will be able to read and write all public and private repository data. This includes the following:
Note: In addition to repository related resources, the repo scope also grants access to manage organization attributes and organization-owned resources including projects, invitations, team memberships and webhooks. This scope also grants the ability to manage projects owned by users.If you don’t like code so much you can just do something else with your life. Have you tried sales or finance?