Federaide is a general purpose multi-agent harness that runs in your terminal. It is meant for recreational programming and automating your scripts. The agents are just named instances of language models which have their own memories, backstories et cetera, and can coordinate with each other as they require. It has its own IDE (complete with interactive structure parsing and jump to def). I am building Federaide as a solo project.
Why do I build this? Some reasons:
- I reached out to some people to try my other IDE, many of them did not have suitable computers (or at all).
- During this time I observed almost every normal person has interacted with AI via their phones (usually the Gemini/ChatGPT app).
- I felt I could bring the joy of coding to a lot more people if I could make an AI assisted IDE run on Android.
- When I go to sleep I like to try out code ideas. Can't take my macbook to bed, have cats. Do have a spare Android, thought it would be nice to have a coding platform I could use in bed.
- Wanted to "vibecode" something, at the time gemini-cli was available, so this was the chosen project (tried a lot of concepts/variants, in rust go etc). Ultimately chose textual over a very simple reason: horizontal scrollbox (absolutely needed that for code display, may be OCD but linewrap does not do it for me).
- Wanted to create a truly powerful AI agent system without worrying about safety. Complete unabashed power was desired.
- Therefore termux native operation was a huge design goal.
- Kept wanting new features, kept adding them too.
- When I started this (in March) multi-agent harnesses were not a thing (arguably they still aren't mainstream, but there are other projects out there now). Wanted to see how different LLMs would react to each other in the same workspace.
- So here we are. It is at an early stage, there are often bugs, but if you find them, I will fix them. Thought I'd share it, it can be daily-driven at this point.
Does it solve a burning problem? Not really. Atleast not one other harnesses don't already (other than maybe providing a performant competent IDE for termux). Is it fun? Fuck yeah. Watch your agents delete the workspace while you sleep. And the excuses are gold :D
Well okay I do want to automate parts of wellsite data processing, the multimodal component of the active skills system is a step towards that, I mean what is geology if not hoarded knowledge and image processing on wetware? But the current models are so far not great at this stuff (I thought I would demo it, but so far the performance has not been up to my standards. Oh well will keep trying.)
This is completely free of MCP. Or any other borrowed compute as much as possible. Except the LLMs ofcourse, though I have tested with Llama.cpp. Everything that can be done on device will be done on device. No logins or email ids or anything like that (unless you count the ChatGPT subscription Oauth, which is not compulsory).
What do I use it for? I mostly use the research tool. Trying the paper trading (active skill) out has been fun (it was made on request for a friend). I like to experiment with orchestration, trying new ideas about memory and stuff like that, with the long term goal of making LLMs learn things outside of fine tuning or direct training. Running go/rust/python code directly on a phone has been fun.
Docs: https://federaide.rocklab.in/manual.html
Github: https://github.com/ROCK-LAB-PRIVATE-LIMITED/federaide
I have not put technical details here for brevity. Feel free to discuss them if you like.