I posted this in the marketplace too, but I'm also posting it here in the AI community!
I wanted to start with something like OpenCore, but the OpenCore settings were too difficult, using API tokens felt too burdensome, and when I actually installed it, it was extremely slow. So I wanted to find a solution within my subscription, and I found takoPI - a project that lets you use terminal agents (like claude-code) via Telegram. I forked it to use my main Mattermost, but the project grew quite a bit. It's still at a proto level, but as I connected Slack and Mattermost (tunaPi), I started thinking: if I'm just chatting with AI (or rather, asking questions about anything), do I really need Slack or Mattermost? So I started creating my own app (tunaDish).
Compared to similar other services, it doesn't have any outstanding advantages (it has more disadvantages actually), but after using it solo and lying in bed with a messenger where I could do almost anything, I got greedy. So I'm working on strengthening branch functionality (it's like thread functionality but in a tree structure), discussion functionality, and multi-agent utilization (currently supporting Codex, Gemini CLI, Claude Code, and OPENCODE).
First, people who dislike it will hate the vibe(?)-coded style. (Honestly, the vibe part only involved chatting through messenger while lying in bed and having music on 😅. For each feature and logic implementation, I had to gather AIs for discussion, consolidate opinions, verify, and get approval 🤯) So it's not really vibe-coding, but rather a strange direction of micro-managing (which was an awkward experience).
The ultimate goal is to simply discuss with AI agents, consolidate the results to create documents, and execute procedures to implement something based on those documents - a kind of IDE (since no one will see it or use it anyway, I'll insist on calling it grandly as an Agent Orchestration Client). To speak more plausibly, the goal is a human intelligence-led development application that builds services based on domain knowledge.
In reality, tunaDish (AOC) was written more than 50% while lying in bed through Mattermost with natural language chatting (do this, do that, this doesn't work, etc.). The rest was done sitting down because desktop app debugging can't be done lying in bed... 😂 I also wrote a lot while traveling. Honestly, there's no major difference from other agent assistant apps (actually, it's pretty much copied 😅 Maybe because the direction is different?).
People who might be interested would probably be:
Want to squeeze value from subscriptions rather than APIs
Have a local LLM but don't want to use it in the terminal, and don't like openwebUI
Terminal agents are the best for coding
Let's create code comfortably while lying in bed!
Those kinds of people, I suppose 😄
The usage is simple. Install it, use Mattermost, Slack, or Telegram to have your subscribed agents discuss while lying in bed or on the couch, and when good ideas come up, ask them to document and consolidate them, then ask them to implement according to the documentation. That's it. Of course, feedback on the results is even better (Sonnet-level is quite good) 😁 Even though it's not at alpha level yet, please visit my GitHub! When you do, please give it a star! 😄 Working on this made me realize what open source is. I found features here and there and added them, and MIT license is the best 😄 Comments with opinions are also welcome 😁😁😁 And tunapi is made to a level that you can fork it (while lying down). Current code tests: 1,023 with 79% coverage (continuing to increase with a 85% goal). If you mainly use Discord or have another chat app (KakaoTalk doesn't work, there's a way but it's too weak!), you can fork it and use it as tunaDish. That's what I'm making this app for anyway! 😄
Thank you! PEACE ☮️🕊️
https://github.com/hang-in/tunaPi
https://github.com/hang-in/tunaDish
