Diff Forge AI

Thumbnail for our Open Source Agentic Development Environement
Using open source TSCircuit, you can do PCB Design using AI Agents
Out of the Box AI Video Editor, ask AI Agents to Edit Videos for you or generate AI Images and Videos
Native Application. Auto Updater from the opensource public github repo releases.
If you have a paid plan, you can register your device to our cloud and sync
You can even setup Twilio to call GPT-Realtime 2 to communicate with your devices and AI Agents via Cellular
Futhermore, if users do have the paid plan, they can control their AI Agents via the Web! The Diff Forge AI App on the Play Store also works

In the Name of God, the Most Beneficent, the Most Merciful.

Inspiration

I am a Canadian and in end of February, I was in Medical School in Tehran, Iran with my little sister, when the war started. Even though I am a Canadian Citizen, they were not able to help us get out of Iran, so we went to the Pakistani Embassy, which organized buses out of Iran. 2 Days to reach the Taftan border between Iran and Pakistan, 2 days at the border, and a 16hr bus drive from Taftan to Quetta which is right under Afghanistan. It was very stressful. We eventually made it to Karachi Pakistan thanks to family members there and came back to Canada. When me and my little sister came back, I went back into making the most of my time by working on startups, and realized the landscaped had changed from copy pasting from Chatgpt to using Codex and Claude Code. This is where I learned about "Bridgemind" and the CEO, Matt, who not only is building the biggest vibe coding community in the world, but also was extremely supportive of his community members, including me. He streams on youtube and is making an Agentic Development Environment which inspired me to make my own, but open source.

I streamed for a few days while vibe coding for a bit (before I decided to lock in) and used GPT-5.5 as my daily driver which wrote over 85% of the code in my codebase, then after the latest models were released, I used Fable 5 to orchestrate GPT5.5, then GPT5.6 Sol Subagents and even let then run 8hr+ runs overnight as I sleep to build this amazing open source piece of software. I have three 20x Codex subs and three 20x Fable 5 subs, and am well known as a “Tokenmaxxer”.

What it does

Diff Forge AI is an open source Agentic Development Environment where developers can develop software, do AI assisted PCB Design, and even AI Video Editing. It has scaffolding for docs, skills, as well as session history so users can launch/return to a session with 1 click.

Diff Forge AI is designed to be the Vibe Coding Toolkit for Developers, featuring multi-terminal workspaces, Audio Dictation (local whisper), and Screen Capture Snipping (works on both MacOS and Windows).

The open source client is fully free to use, but on top of that, there is a paid subscription feature for cloud services such as:

Chat with an AI Agent on your devices through the web, via chat UI or direct terminal shell (web sockets)
Add Cellular Support to call your AI Agents managed via Voice Orchestrator as well as Cloud PBX (via Twilio)
Notifications for when AI Agents are done or need user_input_required
Loopspaces: Create cron jobs, web hooks, or manual triggers for specific workflows And many more features (currently I’m working on AI Email, and both inbound and outbound customer management). The goal is that the ADE Diff Forge AI is where you do all your work, and the web/cellular has both remote control capabilities as well as managing communication with your clients.

How we built it

Codex via VS Code Extension (gpt5.5 era). Without it this project would not exist, or would have taken over 5+ years. Yet I was able to do it alone with only a 200$ subscription and it basically 100x’ed my productivity.

Once Fable 5 released, I saw Theo T3s X post about using Fable 5 to orchestrate GPT-5.5 subagents, used his strategy (which was genuinely clever) and was able to get much more done much faster. The real breakthrough however happened when GPT-5.6 Sol released, it was much better at debugging than Fable 5 and caught way more than it, that I don’t even use adversarial Fable 5 subagents when GPT-5.6 Sol is reviewing. But the best thing that came out of GPT-5.6 Sol orchestrated by Fable 5, was the ability for it to do overnight runs while I slept like a baby. Insane how far AI has gotten. In total we used over 90 billion tokens over the past 2 months for the creation of this project.

Our project was made using:

Tauri for the Native Client (rust backend, react frontend)
NextJS for the Web App
Rust for the Cloud Services The Tauri Native Client also uses open source repos, for example, TSCircuit, which is an opensource project for designing PCBs with react, I simply added that to my Panel structure (which handles multiple terminals and non-terminal panels) and now AI Agents can modify a PCB Design right in the ADE. We used Xterm for Terminal rendering, Enigo Crate for Audio Dictation Insertions, and SCAP crate for Screen Capture/Snipping (which I feed to AI Agents when doing UI Work or have errors).

Challenges we ran into

Running out of Tokens. Thank you Tibo for Resets <3

Accomplishments that we're proud of

Over 888k lines of code across 4 codebases over 2 months, the rust-diffforge client itself is more than 550k lines of code now and is still growing. Complex, functional, compiles, pasts all tests, and is genuinely a technological leap to use the ADE directly, remotely or via cellular (which I do regularly so I can prompt while I’m out and about). Coolest thing is the BYOC (Bring Your Own Cloud) which I had Fable 5 and GPT-5.6 Sol setup, it created a cli (diffforge-cli) and basically after putting your cloud providers API key, it will use the client and auto install and authenticate the device to your account, essentially creating a remote agent from which you can do whatever you want with!

What we learned

Biggest thing I learned, was native app development. I mainly did SaaS development, so this was the first time I was building a native app and learned: MacOS is awesome and Windows is a pain to develop on (I use my MacBook Air M4 now). On top of that, learning about background processes, headless, creating a custom CLI for installing Diff Forge AI and launching it on npm, home-brew, chacalatey, and the difficulties of cross-platform support. Without QA and testing, it’s not possible. Thankfully I have a Windows and MacOS device, but I think I’ll need a linux one as well.

What's next for Diff Forge AI

Currently we are moving towards not only managing AI Agents via Web, but also managing your Customers via AI through contact lists, cellular and email through the web. The goal is that Diff Forge AI will have the toolkit for not only developing products, but also helping solopreneurs scale their business with AI efficiently.