Goose (AI) Chase!

If you haven’t heard of Goose yet then you’re in for a treat!

As a software developer, it’s become a professional requirement to use AI. I could write code without it, but I’d be running about 100 times slower than my AI enhanced colleagues if I did.

In the beginning, there were chat bots…

Like everyone, I started by visiting a website with a chat interface that lets you talk to an LLM (ChatGPT, Gemini, Claude etc.). That was pretty cool but it quickly became painful copying/pasting things back and forth between the chat window and whatever I was doing.

For this reason, those AI chat interfaces have found their way into the tools we use for our work (like VS Code or JetBrains Rider in my case)… which is much better.

However, for security reasons, the agents frequently ask you whether they’re allowed to do X or Y. For big, complex problems you give them, that kind of baby sitting quickly becomes tedious. There are ways to grant them permission to do some things but not others by way of white lists 🥱… but who wants to spend their days doing that?

Then there was OpenClaw

Eventually, it becomes apparent that Agents really should be run in their own sand boxed environments. A couple of months ago I bumped into OpenClaw (aka clawdbot), installed it on a VPS where it could have free reign and hooked it up to Telegram. It was like a breath of fresh air! No more hand holding - I could ask the bot to do genuinely complex things and leave it to do its thing. When it was done, it would ping me and, from wherever I happened to be, I could reply on my phone to ask it to do something else (e.g. give it another github issue to research or try to resolve some review feedback).

At that point though, since I didn’t have to sit there giving the bot permission to do stuff all the time, for the most part the AI no longer needed me. Instead of sitting there waiting for it to write code, I now had quite a bit of time on my hands and I could really easily spin up another few bots to work on other stuff in parallel… then jump between them (a bit like managing a crew of employees). However the way you typically communicate with OpenClaw is via Telegram and that’s a single chat channel… It’s possible to spawn subagents and switch between separate sessions in Telegram, but it’s not very elegant.

Managing multiple agents in Claude Desktop

After a bit of head scratching, I decided to run an OSX VM on my M2 MacBook using UTM… and then run Claude Desktop in the VM. This has a really nice interface for managing multiple chat sessions and, when it’s sand-boxed on a VM, you can run it with --dangerously-skip-permissions enabled so it also doesn’t need hand holding. 🎉

Claude Desktop has two big detractors though.

1. Whilst you can install Claude on one machine (the server) and connect to it from another machine (the client), it’s not possible to use --dangerously-skip-permissions in that configuration.
2. Claude is owned by Anthropic, who treat their customers like cattle. They literally won’t tell you what you get in exchange for your money (and can change what they decide to give you after you’ve paid). They’re dicks basically.

So I started looking around for something that might be good enough to replace Claude Desktop.

The (almost) golden Goose!

After trying a couple of different options, I bumped into Goose AI, which instantly looked pretty cool 😎.

Not gonna lie - it’s got a few rough edges 🙈. Sometimes when I create a new session, it looks like it’s got the context of some other session there (unless I restart) and some of the features aren’t documented at all… but I think I’m willing to work through that stuff because there’s some other stuff about it which is just really cool.

• It’s open source… so if the UI is a little bit janky in places or there are gaps in the docs, there’s nothing stopping you from making a PR to make it better (for everyone).

• You can use it with whatever models you want. That includes using oAuth tokens for subscriptions you’ve got (like Codex) where you pay one flat fee each month and can use the model as much as you want. It also means (like OpenClaw) that you can have lots of models configured and switch between them on whatever basis you like.

• Like Claude Desktop, each separate chat is a separate session by default… but it goes way beyond that, allowing you to fork sessions like you might with Pi, and share sessions with other people.

• It also makes it really easy to see what skills are loaded, schedule work for your agents and do loads of other stuff that I haven’t even had the time to explore yet.

There are still a few wrinkles that need ironing out before I could switch from Claude Desktop to Goose fully. But it is a more than credible alternative - so if you’re using Claude and feeling a bit scorched by some of their recent shenanigans, give Goose a try!

I’ve got hope for this one.