What's next for Beans: Beans UI (call for feedback!)

[hmans]

The next version of Beans is getting two big new features:

• A built-in HTTP server that can serve Beans' GraphQL API over HTTP;
• A Web UI that not only allows you to interact with your Beans, but also orchestrate local agents to work on them.

I want to specifically talk about the latter today. It's inspired by tools like Conductor or Mux, but tightly integrated with Beans (plus some other important differences that I will go into below).

This is what it looks like:

First of all, what can you do with it?

• Create, view and modify your repository's Beans
• Spawn workspaces (with their own isolated git worktrees) where you can let your agent work on individual beans (or anything else)
• Review, commit or integrate the agent's work with the click of a button

How is it different from other tools in this category?

• First and foremost, it runs as a web server (and not a Desktop app). You access the UI through your browser. The idea is that you can put this on a VM and access it from all your devices, including your phone; and your agents will keep working when you close the tab.
• It's tightly integrated with Beans. Beans that are marked as todo have a little "Start Work" button that creates a workspace for them and instructs the agent to start working on the bean right away. Workspaces that modify or create beans automatically list these in their sidebar entry so you always know what a workspace is focusing on.
• Just a single agent chat per workspace. This is a design decision. Other tools usually give you as many chats as you want, which creates its own issues. Beans UI decides to trust the agent's compaction (or large enough context window.)
• It's free and Open Source ;-)

Some limitations and considerations:

• The built-in agent interface currently only supports Claude Code, and the claude executable must be installed (and authenticated) for any of this to work. (So, yes, it uses your existing Claude Code subscription if you have one.) In the long run I want to add support for other popular agents (PRs welcome if you're feeling courageous! But please get in touch first so we can hash out the details first.)
• The current setup assumes you want to integrate (ie. merge) worktrees locally. I know some people would prefer to use GitHub PRs for that, and support for those is planned, but I want to add this in a way that also allows GitLab and other forges.
• The web UI is currently designed to be run from within the repository that you want to work on, and there is no way to interact with multiple projects from within the same UI at the moment. (Since I'm always working on multiple projects in parallel, this will eventually be added, though.)

I'd love for some of you go give it a try and tell me what you think. The best way to run it right now would be to:

• check out this repository's main branch
• run mise run setup to setup dependencies etc.
• run mise run install to build the executables and install them in your PATH

Now in addition to the beans executable you should also have a beans-serve executable. I recommend re-running beans init in your project (and checking/fixing the changes it made to your beans config file); then, running beans-serve will give you the web UI.

Please keep in mind that Beans UI is moving very fast at the moment, with many commits landing on main, so if you're interested in keeping up with the latest changes, don't forget to occasionally pull and rebuild.

Happy to hear feedback if you have any!

[bjesuiter]

This looks extremely interesting to me! I'd want to use it, but since I've fully switched to openai models now, i need the codex or pi.dev support.
Would be willing though to give it a shot and implement it myself in copilot mode with clankers :)

My first idea would be to integrate https://pi.dev since:

• it's minimal
• very extensible
• very hackable
• the base of openclaw, which proves that it's more than capable to be instrumented from the outside.

Just for completeness sake, we could also look into integrating opencode (since it has a webserver approach already / an ACP mode)
oder codex directly (though probably not as standard, not sure if they have an ACP mode right now)

• Update: There seems to be this adapter for the zed IDE for codex-cli, maybe this can work as inspiration.
  https://github.com/zed-industries/codex-acp

Wdyt?

[hmans]

I don't know much about ACP to really know. I was looking at it because I thought it would be a good abstraction layer between Beans UI and agents, but apparently it assumes that there needs to be an extra server running for abstracting an agent, so next to the Beans UI server... there would have been additional servers? That didn't feel like the right direction for me, so I dropped it, but maybe someone with more hands-on ACP experience can weigh in.

ACP or not, support for additional agents would require some sort of abstraction layer that currently doesn't exist (100% of agent code is Claude Code-specific), so it's probably quite a big undertaking. If you and your robot want to have a stab at it, go nuts, I'd love to look at some (working :b) PRs for this!

[bjesuiter]

the thing is, ACP does not need to run as a webserver, but can also simply run as a subprocess, controlled via strings over stdin and stdout. (at least opencode does it this way).

I understand that we'd need some abstraction between, good to know!
So I'll have a look at it and see if i can come up with something working :D

[hmans]

Ah! This is how you know how oblivious I am to some of these new toys :D

Would love to see what you can whip up, and if it can piggyback on ACP, even better.

Always available for a call if you want to do some rubberducking!

[bjesuiter]

I'll take that "always available" literaly! :D Wrote you on bluesky ;)

[andrioidApacta]

Here's some feedback

Agent chat

• The multiple-choice questions can loop. I've found it's much better to answer directly in the chat.
• It's unexpected that when I answer one question, the other two disappear

Planning

• As a new user, my biggest struggle between planning/action was that I had to remember to commit my beans and push them to the remote.
• Alternatively, let the planning always reference main
• I messed up the default branch for my test-repo so that could be the reason.

Overall

• The UI is nice! Already nicer than dealing with Claude Code directly
• Love the worktree support and being able to plan while the agents work
• Can I run /skillname in the agent chat?
• A way to manage skills would be nice
• Some way to preview on the workspaces. Not that running npm run dev and trying it was hard.
• I'd like to be able to see what the sub-agents do; but I guess that's more of a Claude Code limitation.

[rmalchow]

hi,

looks awesome - and i would really like to try this. however ... weird thing:

-----------------------------------------------------------------------------------------------------------------------------------
node@25.8.1     download node-v25.8.1-linux-x64.tar.gz                                                                                               119 B 4m1s [>                 ] ◠
mise ERROR Failed to install core:node@latest: gpg exited with non-zero status: exit code 2
mise ERROR Run with --verbose or MISE_VERBOSE=1 for more information
-----------------------------------------------------------------------------------------------------------------------------------

why does the signature fail?

.rm

[stevegoulet]

Wow that looks amazing. Is anyone else using Beans with the Atlassian MCP to synchronize Jira tickets with Beans? It works great for me so far, but I have to remember to ask Claude Code to update Jira with every change I make to a bean.