ros.ai

Inspiration

Agents can already reach your docs, data and tools, but all of it stays inside their chat. You cannot see what they used or how it connects, so you read transcripts and guess. Every context layer today serves the model. We wanted the one the person can see.

What it does

ros.ai is the open workspace that holds your whole project's context on one canvas. AI agents from any vendor work from it, and you see the whole picture, make the connections and direct them there.

Notes, data, files, code and web research sit on the canvas as blocks. Agents work from them, and results land back with their sources, as proposals you accept or reject. Every block shows which agent made it, on which model, at what cost.

Ros is the built-in assistant: NVIDIA Nemotron 3 running in ros.ai's agent loop with the canvas as its tools. Type one sentence and Ros plans the work, hands tasks to Claude Code, Codex or Gemini CLI, does the web research itself and checks claims.

Ros on this computer is ros-nemotron, our fine-tune of Nemotron 3 Nano 4B, run locally with llama.cpp, free and offline.
Hosted Ros is stock Nemotron 3 Nano 30B and Super 120B on Nebius Token Factory. Sign in and it works, no key needed.

For example: ask "How are Dutch SMEs adopting AI, and what holds them back?" Ros proposes a plan. Claude Code loads the data and runs Python in the sandbox. Codex draws the chart. Ros runs 30 Tavily searches, reads the pages that matter and saves quotes checked word for word. Press Check sources on any answer and each claim comes back Backed, Contradicted or No source. The report cites the canvas, so you see where every number came from.

It runs on your own computer and is open source.

How we built it

TypeScript end to end: React and xyflow for the canvas, Express for the local service, Electron for the Mac app, Python in the macOS sandbox. One agent loop runs any model through the same eighteen typed tools with the same approval checks.

ros-nemotron: Nemotron 3 Super on Token Factory worked 1,351 real tasks through the app as the teacher. 1,140 traces passed the checker and fine-tune Nano 4B with LoRA on one Nebius AI Cloud H100, exported to GGUF for the local runner. Hosted Ros goes through our gateway to Token Factory with a daily allowance per person. We chose the hosted models by test: on ten real agent tasks Super scored 7.5 of 10, Nano 30B 6, Ultra 0.

Challenges we ran into

Small open models write tool calls as text or as JSON in a string. The loop repairs both.
A 4B model has 32k context and our prompt plus tools is 20 KB. Admission now counts tokens, not bytes.
The tool contract froze on 24 October for training while the app kept shipping, so every later integration is additive.

Accomplishments that we're proud of

One canvas, every agent: Claude Code, Codex, Gemini CLI and Ros on one project with one approval queue.
A 4B model that leads: ros-nemotron plans and delegates on a laptop, trained from Super's own work in the app.
Check sources on any agent's answer, quote by quote.
Every result shows what it rests on, every change is undoable, every call and its cost is on the record.

What we learned

A small open model is a good lead when the tools are narrow and typed, not good enough to be the only worker. Once every result shows what it rests on, several agents become usable at once. Measure before you choose; the largest model scored zero.

What's next for ros.ai

Researchers and analysts come first, then product and consulting teams. Next are blocks for designs and meeting notes, so the canvas holds the whole project. Pro is a hosted workspace that keeps running when your laptop is closed; teams get a shared workspace with one rule set and an audit record.