Best of luck with this but I think with so many open source agent managers cropping up, you are going to need to provide very special USP to have people choose yours over the free and open versions.
I guess I would suggest that should be a priority for your site and documentation, to help devs understand what that value offer is.
Your site does seem nicely presented though and clarity in capability is possibly an early win over some of the more chaotic documentation elsewhere.
> Best of luck with this but I think with so many open source agent managers cropping up
What’s the top 5 (or any N) that come to mind:
A) GUI based
B) terminal based
C) web based?
Like, not just personal projects but something with a bit of a community around it? I remember Conductor from a bit ago (seems only Mac is supported) and a few other HN posts but all of those seemed smaller and more barebones. Oh I guess OpenCode also has a desktop and web version, but it never worked well for me (and I need something that can just use headless Claude Code instances).
Asking because I just use Claude Code desktop for organizing my sessions and am a bit behind in that regard - if there are indeed many options that others can vouch for somewhat, I’d love to hear about them!
Edit: apparently there is Cmux (Mac only), T3 Code (very new), Agent Orchestrator (tries to be a weird kanban board), Agor (tries to be a weird canvas board) and Claude Squad (TUI only), but none of those are quite what I'm looking for. If there's all that many options, I might have missed most of them - since Baton or OpenCode (a revisit of it) seem more like what I'd be looking for, maybe Conductor if not Mac only.
I have two claude code subscriptions: a team plan through my employer and I'm paying for the $200/month plan outside of that.
Trading $200/month of my money for the ability to build all of the things I've been thinking about for years is a great trade for me. I've built more things for fun/potential profit in the last year than I did in the previous decade combined.
And of course, one of the things I've built is a version of what OP made that works exactly how I want it to work. :)
It’s so funny to me that every AI user feels the need to add this entire disclaimer about how it’s actually helping them build the Starship Enterprise from scratch or whatever every time someone even hints at it maybe being a little bit of a waste of money.
I build my own products and services and the effective ROI for paying for a more or less unlimited max Claude Code plan is fairly ridiculously positive.
It was more of a tool for myself but some interest from others inspired me so iterating on it. People interested in this kind of thing should join my slack! https://monetworkspace.com/terminal
Everyone is building the things that are just stop gaps: tui/terminal/cli as foundation for UI, worktree instead of jj workspaces. Rust, tauri or python. You all will discover that after the easy fun parts are done you are rebuilding an IDE and at that point why not try to make Vscode better.
- vscode can run as a webapp in a container no electron needed
- vscode supports workspaces, main arguments i hear is vscode assumes one project per window which is just not the case. i have a folder with all my projects/repos that is open as one workspace in vscode
- visualJJ is the best worktree/workspace manager that no one is talking about, i wish it was open source but there is currently no alternative i am aware of
Not many parts are missing to make this perfect for the use cases all the agent managers try to address, but unlike all these hacks this can evolve one extension at a time and i dont have to throw away all the things that my vscode setup has
I don’t know how to phrase this without sounding like an arrogant idiot but seriously: what are people actually programming with agents + worktrees + harnesses + tasks + skills + whatnot? Most workflows I see people adopt involve large amounts of infrastructural fluff only to (more) quickly generate what I (anecdotally) have seen is somewhere between code generation of boilerplatish React/laravel/your-fav-framework components for web or native, and niche toy apps for mostly personal use. My very limited usage of agents has been for scanning large (bloated) codebases to get rid of unused code, meaning time consuming and tedious tasks. But it seems the general trend is that programmers just want faster horses?
You could zoom out a bit and rephrase the question.
Your great-aunt Ida died and left you a consulting team of ten pretty good software engineers. The team's contracts all just ended, so starting tomorrow they'll be idle. Ida said you must run the business for at least two years (fortunately, overhead is already paid for), or forfeit your share of the inheritance. After that you can keep going or liquidate it.
People are just playing around with parallel agents because it looks cool on Twitter. In real prod, 90% of your time isn't spent typing lines of code, it's spent trying to figure out implicit business requirements and debugging undocumented legacy spaghetti. Agents sitting in isolated worktrees are completely useless here - they'll just rapidly and in parallel write code that completely fails to solve the actual business problem
You’re seeing that primarily because it’s what people can show off easily. Side projects they do for fun.
I use all of this stuff daily at work. Normally, I’m working on 2 to 4 features in parallel (so worktrees). This might not be simultaneously, but it’s at least across days or weeks.
Skills, agents, tasks, etc are really about creating repeatability in certain parts of my workflow without needing to be hands on.
Yeah perfect example, the main thing I _would_ use multiple agents on is optimizing/benchmarking code, but for that you specifically can't use worktree, you need one agent per machine or they'll taint each other's benchmarks
Some changes within pre-existing codebases (oh hey, we need feature or mechanism X).
An entirely new internal system with LLM code review, DB migration tracking, time tracking, standups and Teams integration.
A new system that trains neural nets to recognize crops based on Sentinel-2 satellite data (the neural net works okay, mowing and ploughing is harder with a mostly heuristic approach since I don't really have labels).
A new system to migrate somewhere between 1000-2000 forms between proprietary solutions where a team of people have spent a year with limited progress, whereas I'm generating the codegen tool that does most of the work, with the remainder being left up to AI.
A new project linting tool with Go + goja to allow writing rules for validating project stuff in ECMAScript, a bit like ESLint just stack agnostic and can be deployed as a 10 MB executable, to also control stuff like architecture and project conventions that the other tools aren't really geared towards.
Also wrote an OpenAI/Ollama/Claude proxy that allows using on-prem models running on another server through Ollama/llama.cpp and also using AWS Bedrock models when permissions are configured.
Also a bunch of Ansible configuration for stuff like a self-hosted Sentry instance, debugging that piece of shit would be so hard and annoying without something that I can throw logs at (because for some reason they think that having 70 containers running for what should amount to one piece of software is okay).
Also wrote a personal tool that lets me use VLM and Whisper and PySceneDetect and some other stuff to produce EDL so I can take a 3 hour long video and cut it down to 1 hour with LLMs using the transcripts/timestamps (aligned with words, so not too many awkward cuts) that I can then import into DaVinci Resolve for further editing.
Also migrated the apps I host from Contabo VPSes to Proxmox VMs (Hetzner dedicated server from the auction) and went from Docker Swarm + Portainer to pure Docker Compose, also moved from Drone CI to Woodpecker CI and also got rid of the old deprecated Bitnami container images.
Also migrated my homepage from an ancient Ruby and Rails version to more modern ones.
Also wrote a few scripts to replace YOURLS with just an Apache install, the config for which I can automatically append new shortened links to.
I don't even need worktrees or custom skills for most of this, just Claude Code and a subscription, since paying per token would make me go broke.
Please review the site design. Between the thin blue lines appearing & disappearing, and the "television static" in the background I gave up attempting to read anything in the first 30 seconds on the site because my eyes were drawn anywhere other than the content.
Spreading agents across git worktrees sounds awesome right up until the merge step. Sure, they're isolated on the filesystem, but when five parallel Claudes rewrite the exact same base class or interface for their own local needs, you're gonna end up with a merge conflict no neural net could ever untangle. All that saved time will just get burned manually rebasing this parallel chaos. Props for a cool pet project, but conceptually this is an architectural dead end
I started to use superset 2 days ago. Which seems similar. It's pretty nice: https://superset.sh
Fyi: here are some things I would like to have for such a tool
- notification when an agent is done
- each tabs/space has its own terminal, browser, agent
- each tab/space runs in a sandbox (eg docker)
- each tab/space can run my dev server. But must not conflict with the other dev servers running
- each tab/space has a mcp server for the built in browser
Nice to have:
- remote access against my machine/tabs
- being able to make screenshots
I'm confused, I've been running parallel agents on different worktrees within a single view of Claude Desktop for at least a month. I don't see any new features here?
What is your secret sauce, so to speak? I personally built my own local tools and system for this, I tried vibekanban but didn't feel like it added much to my productivity, haven't tried emdash yet.
The main difference is that Baton is agent-agnostic and terminal-native. It doesn't add a GUI on top of Claude Code or Codex, it builds around the terminal itself, so you run whatever agent CLI you want natively, but with convenient shortcuts for launching them. Which is a nicer experience in my view, but people have different views on this.
Baton is also more git-aware. Instead of just showing raw diff line counts, you see commits ahead and behind your target branch, so you can tell at a glance how far each workspace has diverged and shortcuts for resolving it in the matter you want.
One thing I think is unique is the built-in MCP server. It lets agents spawn new workspaces programmatically, so you use an agent to launch agents in new isolated workspaces.
I've built my own as well, in a terminal. Not pretty, but does the job until something better comes along (maybe Baton is that something better): https://github.com/electrovir/agent-storm
This looks dangerously close to cmux but with a narrower focus (Just Claude code)
BTW, the claude app kind supports this with the /remote-control command, and that was what made me move away from cmux (I still have to start the sessions there)
How do you restore the state from the old workspaces? do you spawn tmux and resume the conversation or do you do it differently? from the video it felt like instant
The underlying git worktree still lives on your disk until you delete it. So its not harder than starting a terminal with claude --continue, or codex resume --last inside the git worktree, depending on what agent the user used.
I have tried to provide after best ability, but have only been testing them on vm's on my mac! So be aware. I labeled them Beta due to this. But most features should work fine, probably better on linux than windows.
I have seen teams spend months fine-tuning retrieval algorithms when the real issue was that their ingestion pipeline was feeding HTML boilerplate into the vector store. Fix the input first.
Best of luck with this but I think with so many open source agent managers cropping up, you are going to need to provide very special USP to have people choose yours over the free and open versions.
I guess I would suggest that should be a priority for your site and documentation, to help devs understand what that value offer is.
Your site does seem nicely presented though and clarity in capability is possibly an early win over some of the more chaotic documentation elsewhere.
> Best of luck with this but I think with so many open source agent managers cropping up
What’s the top 5 (or any N) that come to mind:
A) GUI based
B) terminal based
C) web based?
Like, not just personal projects but something with a bit of a community around it? I remember Conductor from a bit ago (seems only Mac is supported) and a few other HN posts but all of those seemed smaller and more barebones. Oh I guess OpenCode also has a desktop and web version, but it never worked well for me (and I need something that can just use headless Claude Code instances).
Asking because I just use Claude Code desktop for organizing my sessions and am a bit behind in that regard - if there are indeed many options that others can vouch for somewhat, I’d love to hear about them!
Edit: apparently there is Cmux (Mac only), T3 Code (very new), Agent Orchestrator (tries to be a weird kanban board), Agor (tries to be a weird canvas board) and Claude Squad (TUI only), but none of those are quite what I'm looking for. If there's all that many options, I might have missed most of them - since Baton or OpenCode (a revisit of it) seem more like what I'd be looking for, maybe Conductor if not Mac only.
Oh-my-Claude and oh-my-codex (same creators) seem to be popular. The latter was used for the immediate ports of the Claude leak to python and rust.
https://github.com/instructkr/claw-code
Here's the oh-my-openagent that is also name checked in the link above:
https://ohmyopenagent.com/
I appreciate the feedback!
How can people afford to use Claude Code like this‽ Is everyone just playing with it on their employer's dime or what?
I have two claude code subscriptions: a team plan through my employer and I'm paying for the $200/month plan outside of that.
Trading $200/month of my money for the ability to build all of the things I've been thinking about for years is a great trade for me. I've built more things for fun/potential profit in the last year than I did in the previous decade combined.
And of course, one of the things I've built is a version of what OP made that works exactly how I want it to work. :)
It’s so funny to me that every AI user feels the need to add this entire disclaimer about how it’s actually helping them build the Starship Enterprise from scratch or whatever every time someone even hints at it maybe being a little bit of a waste of money.
I build my own products and services and the effective ROI for paying for a more or less unlimited max Claude Code plan is fairly ridiculously positive.
Like you make money with them?
This uses the CLIs so its using subscription pricing, not token pricing
VC funding + spending more money on Claude instead of hiring more engineers
200 dollars a month goes a long way with claude code
I started building a similar project for myself, a terminal PTY running through a desktop daemon: https://youtu.be/6KY-HCn3SaA
The fun part being it worked on mobile too: https://youtube.com/shorts/CmemwDGwpx8?si=xzAJBb8ha7DLIDmY
It was more of a tool for myself but some interest from others inspired me so iterating on it. People interested in this kind of thing should join my slack! https://monetworkspace.com/terminal
Everyone is building the things that are just stop gaps: tui/terminal/cli as foundation for UI, worktree instead of jj workspaces. Rust, tauri or python. You all will discover that after the easy fun parts are done you are rebuilding an IDE and at that point why not try to make Vscode better.
- vscode can run as a webapp in a container no electron needed
- vscode supports workspaces, main arguments i hear is vscode assumes one project per window which is just not the case. i have a folder with all my projects/repos that is open as one workspace in vscode
- visualJJ is the best worktree/workspace manager that no one is talking about, i wish it was open source but there is currently no alternative i am aware of
Not many parts are missing to make this perfect for the use cases all the agent managers try to address, but unlike all these hacks this can evolve one extension at a time and i dont have to throw away all the things that my vscode setup has
I don’t know how to phrase this without sounding like an arrogant idiot but seriously: what are people actually programming with agents + worktrees + harnesses + tasks + skills + whatnot? Most workflows I see people adopt involve large amounts of infrastructural fluff only to (more) quickly generate what I (anecdotally) have seen is somewhere between code generation of boilerplatish React/laravel/your-fav-framework components for web or native, and niche toy apps for mostly personal use. My very limited usage of agents has been for scanning large (bloated) codebases to get rid of unused code, meaning time consuming and tedious tasks. But it seems the general trend is that programmers just want faster horses?
You could zoom out a bit and rephrase the question.
Your great-aunt Ida died and left you a consulting team of ten pretty good software engineers. The team's contracts all just ended, so starting tomorrow they'll be idle. Ida said you must run the business for at least two years (fortunately, overhead is already paid for), or forfeit your share of the inheritance. After that you can keep going or liquidate it.
What do you do?
People are just playing around with parallel agents because it looks cool on Twitter. In real prod, 90% of your time isn't spent typing lines of code, it's spent trying to figure out implicit business requirements and debugging undocumented legacy spaghetti. Agents sitting in isolated worktrees are completely useless here - they'll just rapidly and in parallel write code that completely fails to solve the actual business problem
You’re seeing that primarily because it’s what people can show off easily. Side projects they do for fun.
I use all of this stuff daily at work. Normally, I’m working on 2 to 4 features in parallel (so worktrees). This might not be simultaneously, but it’s at least across days or weeks.
Skills, agents, tasks, etc are really about creating repeatability in certain parts of my workflow without needing to be hands on.
Yeah perfect example, the main thing I _would_ use multiple agents on is optimizing/benchmarking code, but for that you specifically can't use worktree, you need one agent per machine or they'll taint each other's benchmarks
Some changes within pre-existing codebases (oh hey, we need feature or mechanism X).
An entirely new internal system with LLM code review, DB migration tracking, time tracking, standups and Teams integration.
A new system that trains neural nets to recognize crops based on Sentinel-2 satellite data (the neural net works okay, mowing and ploughing is harder with a mostly heuristic approach since I don't really have labels).
A new system to migrate somewhere between 1000-2000 forms between proprietary solutions where a team of people have spent a year with limited progress, whereas I'm generating the codegen tool that does most of the work, with the remainder being left up to AI.
A new project linting tool with Go + goja to allow writing rules for validating project stuff in ECMAScript, a bit like ESLint just stack agnostic and can be deployed as a 10 MB executable, to also control stuff like architecture and project conventions that the other tools aren't really geared towards.
Also wrote an OpenAI/Ollama/Claude proxy that allows using on-prem models running on another server through Ollama/llama.cpp and also using AWS Bedrock models when permissions are configured.
Also a bunch of Ansible configuration for stuff like a self-hosted Sentry instance, debugging that piece of shit would be so hard and annoying without something that I can throw logs at (because for some reason they think that having 70 containers running for what should amount to one piece of software is okay).
Also wrote a personal tool that lets me use VLM and Whisper and PySceneDetect and some other stuff to produce EDL so I can take a 3 hour long video and cut it down to 1 hour with LLMs using the transcripts/timestamps (aligned with words, so not too many awkward cuts) that I can then import into DaVinci Resolve for further editing.
Also migrated the apps I host from Contabo VPSes to Proxmox VMs (Hetzner dedicated server from the auction) and went from Docker Swarm + Portainer to pure Docker Compose, also moved from Drone CI to Woodpecker CI and also got rid of the old deprecated Bitnami container images.
Also migrated my homepage from an ancient Ruby and Rails version to more modern ones.
Also wrote a few scripts to replace YOURLS with just an Apache install, the config for which I can automatically append new shortened links to.
I don't even need worktrees or custom skills for most of this, just Claude Code and a subscription, since paying per token would make me go broke.
Please review the site design. Between the thin blue lines appearing & disappearing, and the "television static" in the background I gave up attempting to read anything in the first 30 seconds on the site because my eyes were drawn anywhere other than the content.
Appreciate the feedback! Looking into it
Spreading agents across git worktrees sounds awesome right up until the merge step. Sure, they're isolated on the filesystem, but when five parallel Claudes rewrite the exact same base class or interface for their own local needs, you're gonna end up with a merge conflict no neural net could ever untangle. All that saved time will just get burned manually rebasing this parallel chaos. Props for a cool pet project, but conceptually this is an architectural dead end
[dead]
Very cool. And congrats on the launch.
I started to use superset 2 days ago. Which seems similar. It's pretty nice: https://superset.sh
Fyi: here are some things I would like to have for such a tool - notification when an agent is done - each tabs/space has its own terminal, browser, agent - each tab/space runs in a sandbox (eg docker) - each tab/space can run my dev server. But must not conflict with the other dev servers running - each tab/space has a mcp server for the built in browser
Nice to have: - remote access against my machine/tabs - being able to make screenshots
Nice. My take on this is https://github.com/rcarmo/piclaw and https://github.com/rcarmo/webterm since I prefer to run my agents away from my desktop but still have a nice UX, but I have been thinking of packing them in electrobun.
I'm confused, I've been running parallel agents on different worktrees within a single view of Claude Desktop for at least a month. I don't see any new features here?
Fair but FWIW I love a GUI and I’m not gonna complain if everyone and their mother want to offer options
Let a thousand vibecoded flowers bloom
Nice work! Congrats on the release, did you check out Vibe-Kanban or Emdash which are both building in this space?
https://www.emdash.sh/
https://vibekanban.com/
What is your secret sauce, so to speak? I personally built my own local tools and system for this, I tried vibekanban but didn't feel like it added much to my productivity, haven't tried emdash yet.
Congrats on your launch! How is this different than Conductor?
The main difference is that Baton is agent-agnostic and terminal-native. It doesn't add a GUI on top of Claude Code or Codex, it builds around the terminal itself, so you run whatever agent CLI you want natively, but with convenient shortcuts for launching them. Which is a nicer experience in my view, but people have different views on this.
Baton is also more git-aware. Instead of just showing raw diff line counts, you see commits ahead and behind your target branch, so you can tell at a glance how far each workspace has diverged and shortcuts for resolving it in the matter you want.
One thing I think is unique is the built-in MCP server. It lets agents spawn new workspaces programmatically, so you use an agent to launch agents in new isolated workspaces.
Would be curious if it is more polished than Conductor. Memory leaks and random bugs seem to crop up in Conductor far too often.
If nothing else, I see that Conductor is currently Mac only.
I think every developer have done running multiple Ai agents for one work. I even tried to make them talk over slack :)
I couldn't see the pricing page tho.
Maybe I'm daft, I watched the video, and I just didn't understand what this is, or why I'd use it.
Seems like just tabs of claude code, plus markdown viewer which can just be another tab (with an editor) in a tabbed terminal?
My ide supports multiple terminal tabs, plus is a project aware code viewer, and has the ability to run the project.
What would I gain by using this?
I've built my own as well, in a terminal. Not pretty, but does the job until something better comes along (maybe Baton is that something better): https://github.com/electrovir/agent-storm
This looks dangerously close to cmux but with a narrower focus (Just Claude code)
BTW, the claude app kind supports this with the /remote-control command, and that was what made me move away from cmux (I still have to start the sessions there)
Theo's t3code does a lot of this for free I think. Interested to know if it uses the same trick for accessing Claude without violating their TOS.
https://t3.codes
This looks impressive!
How do you restore the state from the old workspaces? do you spawn tmux and resume the conversation or do you do it differently? from the video it felt like instant
The underlying git worktree still lives on your disk until you delete it. So its not harder than starting a terminal with claude --continue, or codex resume --last inside the git worktree, depending on what agent the user used.
Looks like it could be extremely useful. I'm developing one thing at a time - small projects - but I'll keep this in mind for when things grow!
Nice tool for working multiple sessions without them tripping over each-other.
I appreciate that you provided multiple OS versions rather than just go for Mac only like some.
I have tried to provide after best ability, but have only been testing them on vm's on my mac! So be aware. I labeled them Beta due to this. But most features should work fine, probably better on linux than windows.
I have not done much multi-agent development. Trying to understand what problem this solves, surely one can spin up multiple terminal tabs?
I have not dove into the particulars, i'm assuming the agents do push/pull requests on your repo so no versioning issues.
Are agents at worktree level or can a single agent and chat work on a parent directory above multiple worktrees of different repos?
You can open a directory also as a workspace, it just wont have git stats and git shortcuts.
> Features
It’s blank. Lots of blank gray rectangles too. Site is broken?
Everyone's building the same thing nowadays ^^
This looks great. How do you compare to cmux?
It's quite impressive
Go away, I'm baitin'!
[dead]
[dead]
[dead]
[dead]
[flagged]
[flagged]
I have seen teams spend months fine-tuning retrieval algorithms when the real issue was that their ingestion pipeline was feeding HTML boilerplate into the vector store. Fix the input first.