Journal: Why I am excited for Theia AI

[codingkriggs]

Summary

• I'm from an Eclipse background
• Most AI IDEs and extensions don't empower their users
• Theia has a great design for this problem space
• Architecture-specific prompts, and new architectures that are fundamentally designed to work well with AIs, are in my view where things will inevitably end up.

My Background

I am a long-time supporter of the Eclipse foundation, Java-based Eclipse IDE, and Eclipse RCP and PDE. Also EMF and CDO. I also have worked with Eclipse Che over several years, and through that project have worked with Theia.

My Theia Story

What has excited me of late, and brought my focus heavily into Theia, is the announcement of Theia AI. I know there are a couple dozen IDEs and extension now that are all making an AI play, but they're all doing the same thing, taking away the responsibility of how the IDE interacts with the AI with some secret sauce algorithm. Whoever comes up with the best secret sauce wins.

Another way to describe the "responsibility" they are taking away is "empowerment". Generally a hallmark of open source, tools that are open source give a lot of control to power users, with sensible defaults. And Theia is actively approaching this with an open source philosophy, and that is its strength.

Deciding what system prompt is fed into the AI along with the user's prompt will be essential to future software architectures, I believe. I don't agree with the "give AI all your code, plus 2-3 words from the user, and have the IDE fully control the system prompt, and watch it infer everything and give you a result" philosophy. Especially in the short term when many code bases do not fully fit in the AI context. I don't trust any automatic mechanism that grabs files without the user's input: it alternates between eating up all my context and all my credits by giving the AI too much code, and then giving a poor answer because the context is all used up, or not giving enough code, which also results in a poor answer.

A domain- and architecture-specific system prompt will give better results. Sure you could remind the AI to consult a certain file as documentation before giving a response every time you talk to it, but this is an imperfect band-aid. Mapping each prompting agent to a most suitable model, giving AI functions appropriate to the task, saving the developer from writing the same preamble every prompt, knowing that the platform is not, without the user's say, telling the AI to do something that may prematurely use up their context, this is a step beyond polish. It makes the tool actually usable and helpful.

I am convinced that not only will our tools be reshaped by AI to help with our software architectures, but that future architectures will show themselves to be fundamentally reshaped to play better with AI, in other words, "meeting in the middle". This transformation could be as big as the contrast between looking at code designed for a monolithic architecture and code designed for microservices. This is my project's primary thesis right now, and Theia is the tool to help me get there.

[JonasHelming]

Absolutely love this journal — it describes our common vision for Theia AI perfectly!

We're especially happy to see that:
a) we were able to communicate this vision clearly enough for others to understand and resonate with it, and
b) that community members like you not only share the vision, but arrive at these conclusions independently based on real needs and insights.

Your articulation of empowerment vs. hidden magic, and the importance of domain- and architecture-specific prompts, mirrors exactly why we built Theia AI as a flexible framework, not just another black-box AI feature set.

We'd love to see more posts like this — thoughts, questions, feedback, even challenges! And if you're ever interested in contributing, whether it's via code, docs, or just ideas, the door is always open.

If you choose to share these reflections on social media, please tag us — we’ll happily amplify the message. The community benefits when more people see how Theia AI is different, and why that matters.

Thanks again for this thoughtful and inspiring post!

[codingkriggs]

Here's a timeline of events:

• First heard of Theia AI in its official announcement on EclipseSource several months ago.
• Made a note to check it out "one day".
• With every passing week, I'd been increasingly seeing AI being a integral part of software development.
• A turning point for me was Claude 3.7, and soon after, trying out Claude Code, where it produced large amounts of code and did it well the first time.
• Figured I needed to catch up on the millions of AI IDEs out there (keeping up with AI is like a full-time job), and in parallel, was doing some research related to Eclipse Che.
• When I was reading stuff about RedHat's OpenShift and okd, I saw a mention of Theia and remembered there was that Theia AI I hadn't yet checked out, and I was now more interested than ever in an AI IDE solution.
• It helped that, through Che, I already had experience with Theia and also its extension infrastructure and how it differs and overlaps with vscode.
• Looking for recent videos about Theia AI, I saw a link to your Mozilla Builder talk: https://www.youtube.com/watch?v=Cz9Rh9jVHz4
• By the way, superb presentation, it was very engaging for its whole length, and I've watched it several times, so kudos.
• But for me a few wow moments were when the "AI Agent History" was open and I was excited that I got to see exactly what was sent out, not just my user prompt, followed soon by when you showed how you could edit that prompt. That's when Theia won me over.

Plus you played to my love of open source architectures and ecosystems and my career-long history with technologies under the Eclipse umbrella (not just IDEs but many of the technology projects as well).

Anyway, given that I watched that video, it is no accident that what I wrote mirrors a lot of what you said in it.