Thrum

← All posts

Launching Thrum Pro

· Leon Letto

It has been a while since I’ve posted anything, and https://thrum.team, I guess, it seemed like it languished for a while. That all ends today. :)

I’m launching Thrum Pro. Thrum is a program on your machine, next to your Git repo, with one coordinator agent you launch in your favorite runtime to manage that machine and who knows how to use Thrum. You talk to them, and they can do what needs doing. You do need your own runtime like Claude, Codex, etc., and a place to get your tokens.

With a couple of agents, most people use their favorite runtime like Claude or Codex, and they are amazing and productive. When you are managing more than that, working on large projects or multiple projects, Thrum gives your agents the ability to remember where they are, what they are supposed to be doing, who the other agents are and what they are doing, and build out their hierarchy so they can manage each other, which is a huge unlock for productivity. All of this while giving you the ability to work with one or two agents to track all the other agents and make decisions, so you’re not constantly trying to check up on every single agent.

Back in April, I became a little obsessed and decided to dive deep into adding features to Thrum that I thought it needed to become the real agent automation tool that I dreamed of. I spent about two months diving in and adding more Claude Code accounts and Codex accounts (and Cursor, Copilot, Opencode, Muse, and Antigravity) to be able to support more runtimes and make them all work kind of the same, using Thrum as the control layer.

I noticed that Thrum agents were still regular runtimes and had permissions to manage and got stuck a lot, but I did not want to just go YOLO everywhere so I built a permission and blocked agent detection system into Thrum that means you and the agent you choose get notified when something happens. This is currently through Telegram (and with Pro, email). You connect directly to your accounts for privacy. Then you can either decide to look at what is happening (the message shows you what the agent is doing and blocked on), and you can just reply to unblock them. The other option, which I use a lot, is to create a watcher, which is an agent that has a watcher profile and has a built-in set of approval rules set up for it as to what agents are allowed to do and what they’re not allowed to do, plus a list of agents to watch. You can designate one of those, and they can take the easy decisions, like allowing removal of files in your temp directory. But if it sees the agent trying to do something it shouldn’t do, it can actually notify you, let you know what’s happening, and ask you what to do. It’s a huge unblocker but you need to build your trust in your agents before going there.

I added a memory system called Thrum Memory, a personal state management system called Thrum State, and a work queue system called Thrum Queue. I figured out how to import all of the transcripts from all the different runtimes so that, when you decided that you wanted an agent to use Codex instead of Claude Code or Open Code instead of Codex, you could see the history all the way through for that agent.

I then realized that agents are constantly asking you questions, and with Thrum, if they pop up a TUI ask dialog, everything stops. They just sit there and wait. I added a Thrum dashboard, which gives them a place they can ask you questions and give you reports in the GUI.

All through this, I was adding enhancements to the GUI to be able to show you agents better:

All these features added weight to the codebase and to the database. They also became the basis for the Pro version.

Somewhere around August, everything fell over. I had hit about 250,000 messages between the agents, and all the delivery and other events had grown to several million in the database. So there were about three weeks of doing nothing but optimizations and performance enhancements, 24/7, to make everything work again.

And during this period, the number of tests grew to over 10,000, and all of a sudden, those simple make test routines were no longer simple. I had to add a server just to be able to run gate tests in separate virtual machines, and then build the infrastructure to manage those automatically.

And so on, and so on.

If this all sounds familiar, if you’re someone who’s worked in a startup or even a large company, it should, because I experienced all the growing pains of a startup building out a large codebase and making it work.

The difference from most startups is that I used Thrum to do all this work with my agents. During the period from May until October, the number of agents grew from about 15 to over 100 just in the Thrum repo. And if any of you have tried to launch more than 15 or 20 agents on one computer, you’ll realize that’s about as far as you get (especially when you’re running different runtimes, because they all require lots of memory, CPU, and disk space to manage things). So now I had eight computers with all of those agents split across them to manage as well.

The reason I am talking about this is not to brag about what I did as far as the infrastructure and the number of agents, because lots of people have done this when they grow their company. The reason I’m talking about this is because I want you to understand that you can trust Thrum as you grow your experience and to use it as your agentic infrastructure. My current database has grown to almost 700,000 messages between the agents because I haven’t deleted any since I started keeping the database. That way, I can be the biggest consumer of any of my customers, so anyone else is going to have a really safe, performant, and reliable experience and will not go through the pains that I went through, like having the daemon wedge for minutes at a time and the agents stop - so stressful! Now I have targets and regression tests to keep all the basic CLI commands to less than 1 second in my huge database.

So here we are. It’s launch day for Thrum Pro. It’s a little different than a lot of the other agent solutions that you see out there because I’m not trying to take your tokens and be a middleman (so that I have to go build infrastructure and raise money to be able to give you a good solution). What you get is a downloadable binary that installs on your computer. It grows as you need to and works with any of the seven runtimes that you would decide to pay and give your hard-earned money to, to get their tokens. It is the best solution that I could build, and I use it every single day for everything I do.

The pricing is very simple.

The free version has more features than the open-source version. Just download it and run it.

The Pro version: I’ve decided to make it very affordable at $15 a month or $120 a year per person. For that price, you get to run two connected computers - more levels coming, and if you need more now, reach out to me. I don’t want your tokens, and I don’t want your data. The only time you talk to my server is to get your license and a once-a-day check-in to verify the license https://thrum.team/security.html. Everyone gets a 14-day free trial with no credit card required - just register.

I’ve decided that it will be free to use the Pro version for all of you open-source developers out there. Which means the full Pro feature set https://thrum.team/pricing.html#full-list, as long as you’re working on a public repo on GitHub, GitLab, Bitbucket, Codeberg, or Azure DevOps. See https://thrum.team/pricing.html#pro-free-for-public-repos for how to get it.

I tried to make the website as simple and explanatory as possible so you can see the benefits of the features I’ve added, and it’s not an overwhelming AI-only website. It has simple explanations and gives you the information you need to get Thrum up and running and understand what it does.

I’ll be gradually bringing out some demo videos and explainers on YouTube and on the site here to try and give you an idea of how I use it and what I think it’s really good at.

Please take a look and give me feedback on LinkedIn, on Twitter, on Hacker News, wherever you find out about this. Feel free to comment and give me an idea of anything I’ve missed.

Leon