A native macOS app that manages persistent multi-AI collaboration scenarios — resume anytime, fully isolated.
Why We Built This
Over the past year, our team has been using AI coding tools to conduct on-device continual learning research. At first, one terminal window running a single Claude or Codex session was enough. But as our team grew and research complexity increased, we ran into four problems.
Onboarding Takes Too Long
New team members — including colleagues overseas — spent days understanding the project environment and configuring toolchains before doing any real research. We had accumulated significant digital intellectual assets but lacked a fast entry point to reuse them.
Parallel Experiments Are Too Fragile
Talented researchers constantly needed to run two or three experiments simultaneously with different parameters and hypotheses. But doing this in a single environment was extremely brittle — directory conflicts, process cross-talk, state pollution. Any one of these could destroy days of work.
A Single Model Cannot Do Everything Well
Experience taught us repeatedly: you cannot let the same model both write code and review it — not even different versions of the same model. The most effective combination is Codex for execution and Claude for review — multi-model collaboration is how you get things right. You can even bring in locally hosted models like Hermes, letting multiple high-quality models collaborate as agents together.
Interrupted Experiments Mean Lost Progress
A research experiment often spans multiple days. Shutting down, switching tasks, or even just restarting a terminal — the conversation context, work progress, and environment state are all gone. We needed to pause and resume at will, making state management invisible rather than burdensome.
These four problems are really one problem: we lacked a local infrastructure genuinely designed for multi-AI collaboration. Existing tools are either cloud-based (data must be uploaded), single-agent (no coordination between AIs), ephemeral (close the window, everything resets), or vendor-locked with cumbersome integration. None solve isolation, collaboration, and persistence together.
So we built one ourselves.
What AI Collab Is
AI Collab is a native macOS application and a command-line automation tool. In one sentence: it organizes your local AI coding tools into a team with memory, collaboration, and recoverability.
Its core concept is the Scenario — a persistent work room with its own isolated code workspace, independent participants, independent collaboration routing, and independent lifecycle. You can run ten scenarios simultaneously, each running different experiments, completely isolated from one another.
A Real Use Case
You are researching a model optimization approach. You create a scenario called "research-kv-cache-policy" with two participants: a Codex instance as the analyst, writing code and running experiments; a Claude instance as the reviewer, checking code and validating conclusions. They exchange review requests and feedback through a typed message system. You leave for the day and close the scenario. The next morning, you resume with one click — Codex picks up yesterday's conversation, and Claude remembers where it left off in the review.
Multi-Model Collaboration: Not a Gimmick, a Necessity
The deepest lesson we learned over the past year: letting one AI write its own code and review its own code always results in self-deception — especially when different models from the same model family are fundamentally constrained by the same RL training ceiling.
You should not hand a task to a single expert. Instead, you should let multiple experts collaborate to get it done.
AI Collab lets you configure multiple participants in a scenario, each an independent AI instance. You choose different models, different roles, different permissions. Through collaboration policies, you define who can message whom, which message types require replies, and which are notifications only.
Works With Any Project
AI Collab is not tied to any specific project. It understands your project structure through a generic adapter protocol: you place two configuration files in your project root describing your repositories, their organization, and environment setup.
If your project directory does not have these files yet — no problem. When you register a project directory in the App, AI Collab automatically scans the directory structure and generates configuration drafts. Confirm them and start creating scenarios immediately.
We are our own first users. The first project AI Collab manages is our own EdgeStudio — a fully-featured Harness environment and an on-device AI research platform spanning over a dozen repositories. We use AI Collab daily to create scenarios, run experiments, and do code reviews. The code changes mentioned in this article were completed inside AI Collab scenarios.
Open Source Today
AI Collab is open source today under the MIT license. You can:
- Download the dmg and install directly
- Build from source
- Write your own project adapters
- Contribute code
We do not think multi-AI collaboration is a niche need. As AI coding tools grow more powerful, the "one person, one AI assistant" model will rapidly evolve into "one person directing an AI team" (and no, we are not talking about OPC). That team needs isolated workspaces, reliable communication, and recoverable state management. These are exactly what AI Collab provides.
If you are using multiple AI tools for development or research, if you are tired of reconfiguring environments every restart, if you want different AIs to truly coordinate instead of working in silos — give it a try, send us feedback, and help us build it together.