Devlog 8

This version is mostly Quality of Life improvements and stability, but it really improves a lot.
This update brings:

Backend

• Working config.json file for basic model configuration
• Not crashing every second prompt
• webfetch/websearch somewhat faster and not crashing anymore
• Better security
• Improved SearXNG default configuration

Frontend

• Mobile interface — the webpage finally works on phones
• Regeneration — now, even when the backend throws an error, the frontend gives you the option to retry the request
• Reasoning popup — you can now see what the model is actually thinking about

WIP

• File attachments — frontend part finished, not wired into the backend

I’m probably going to stop working on new features and rather start finishing this project up, so probably another QoL update will be next, alongside one last large feature I need to do: proper demo account creation, ’cause right now demo mode is just one large shared account. I know, pretty bad.

devlog 7

In this version I finally added full MCP support. Websearch also got a few improvements, plus the full Lightpanda browser through MCP integration. Backend has improved stability.
I have started on the mobile interface, although it’s not yet ready.
I’ve also started working on the config system, for now most of it is in .env, but soon I’m gonna replace it with a proper .yaml file.

Devlog 6

The demo is finally out on slop.racek.xyz.

The demo won’t have all the features of the full dev build, because my server isn’t able to run proper embedding and reranking models. That’s why the demo relies on the Tavily API for web searches, it’s nearly identical.

The harness is finally starting to come together. I implemented just two tools: websearch and webfetch. This took me so long because I wasn’t only adding those tools, but also benchmarking different embedding and reranking models, so I can be sure I’m not losing performance for no reason.

The websearch tool is definitely going to go through more improvements, but for now this is the version I’m okay with and the AI also seems to like it (yes I asked it)

Devlog 5

After some time, I’m back with new features and improvements. I finally added a Reasoning/Websearch indicator so users can see what the AI is currently doing. This gives users a sense that something is happening, which makes the entire process feel faster.

I’ve also removed the search button from the sidebar since I’m not going to work on it anytime soon. I’ll probably implement it for the second release, but not right now.

If everything goes correctly, today I hope to release a demo with DeepSeek V4 Flash as a free model. All I need to do before releasing the demo is:

1. Finish demo sign-in
2. Fix a few frontend bugs
3. Add rate limiting to the backend
4. Deploy on my server

After releasing the demo, I’ll have a really basic AI chat interface: you send a message, the AI responds. It has one tool, websearch. That’s why after releasing the demo, I want to focus on the harness and make it actually good.

I’m definitely going to do a SearXNG implementation for the websearch tool (currently it uses the Tavily API) and implement a custom page loader/parser and some basic RAG. I haven’t used RAG for a long time, so I’m going to have to do some research on it.

I definitely want to add more tools later on, but I’ll keep that for future devlogs.

Devlog 4

Holy moly, it’s again another 9 hours. Thankfully, I finally have something to talk about in this devlog.

I’m getting really close to publishing the demo page. I have a mostly working frontend, and the backend has just a few bugs I need to handle. I’m still missing rate limiting and model verification on the backend. For now, the harness has just one websearch tool. in the future, I want to add much more and make it a full harness, but for the first release I hope this is enough. One last feature I want to implement on the frontend is displaying the harness state, so instead of showing the user a simple loader, they will see what’s currently happening in the harness — what the agent is reasoning about, and so on.

So the plan for now is:

1. Complete frontend integration
2. Fix backend bugs
3. Display harness state on frontend
4. Release public demo page with rate limiting
5. Improve harness
6. Ship first release

Devlog 3

After a week I’m finally back home and continuing my work on this
project.
I noticed I’ve already gone 9 hours without a devlog, but I don’t have
much to talk about yet. Well, I started working on the backend — it’s
taking quite slow because I’m giving it the proper time to ensure its
security and ease of implementation for the frontend. I guess I’ll just
list the stuff I have implemented or am working on implementing:

• JWT Authentication (finished)
• Conversation Branching (finished)
• AI Harness (mostly finished)
• DB implementation (mostly implemented)
• Rate limiting (not yet)
• API
• User focused / Authentication (finished)
• Inference / AI (not started yet)
  Still got some work to do, but the backend is getting close to finished.
  The last step will be frontend API implementation, and the project will
  be ready for demo deployment.
  I’ll probably launch the public demo without the need to create an
  account, with free inference on DeepSeek v4 Flash. I’ll use the OpenCode
  Zen API for that because they offer free DeepSeek v4 Flash inference. I
  hate the person who tought is good idea to block publishing copied
  messages i didn’t use AI to write it but i wrote it in my code editor
  and now i have to rewrite this entire message cause stardance thinks i
  AI generated this.

Devlog 2

The frontend is finally in a usable state, which means I can finally start working on the backend.

I intentionally kept the frontend simple no special features yet. I have plenty of ideas for special features and improvements, but my first priority is getting the application into a polished, presentable state. Once that’s done, I’ll start experimenting with those ideas.

Now, let’s talk about the backend. What do I actually want?

I’ll probably use basic Python with SQLite. There’s no need for a separate database for this project. The backend will include simple JWT authentication, response streaming, model selection, and some basic rate limiting.

Once everything is working reliably, I’ll start thinking about building a custom harness or maybe even experimenting with some multi-agent workflows. But first, I don’t want to get sidetracked, I want to finish the base application.

Devlog #1

I’m starting with a simple agentic web interface.

For the first release, I want to focus mainly on the user interface and finishing the frontend. The backend will just be a simple API or an Ollama proxy with rate limiting.

I also want to create a custom harness in Python, but not for the first release. I prepared some design files beforehand, but they look bad, so I’m not really following them much.

If I don’t mess up the UI, it should come out looking okay, but I’m a bit more worried about the UX. I like putting animations everywhere, which always absolutely ruins UX, so I’ll try to be more conservative with them this time.

I’m working on the frontend mostly without AI, using it only as interactive documentation. I’m hoping to get the frontend into a usable form asap so I can start working on other features and the custom harness.