Direct AI agents on the context you can see.

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

Get ros.ai Try the live demo
Claude Code
Codex
Gemini CLI
Cursor
Pi
Hermes Agent
Goose
OpenCode
Ollama
LM Studio
ROS
Your own
Agents
Your whole project
Sources
Notion
Google Drive
GitHub
Web search
Linear
Google Docs
Google Sheets
arXiv
Zotero
Overleaf
Files and data
Python sandbox
Local ROS on Nemotron 3 Nano2 agents working
Notion pageTraining plan14 steps · read live
Fileannotation_guide.pdf12 pages
Dataros_sessions.jsonl12,400 turnsGemini CLI
Group · Evals
Text · FindingTool calls right 84%, up from 71%3 backedClaude Code
Chartval_loss.pngfrom train.logCodex
Notebookfinetune_lora.ipynb
Ask ROS, or @ an agent1 to review

Agents work from your context, with your tools.

Your notes, data, files, pages and code sit on one canvas. Agents from any vendor read it, use the tools you connect and put their work back on it.

  1. 1

    Your context

    Notion pageTraining plan
    Dataros_sessions.jsonl
    PDFNemotron 3 report
    Codetrain_lora.py

    One canvas. Everything the work touches.

  2. 2

    An agent reads it

    Claude Code Codex Gemini CLI CCursor Local model ROS

    Any vendor. Your plan or your key.

  3. 3

    It uses your tools

    Web search Python sandbox Notion Google Drive GitHub Zotero

    Every agent gets the same tools.

  4. 4

    The work lands on the canvas

    Text · Finding Tool calls right 84%, up from 71% 3 backedClaude Code

    With its source, its agent and its model.

See the whole picture.

Every source, search, run and answer is a block on the canvas. Each one shows where it came from, which agent made it and on which model.

WebLoRA on small models24 searches · 4 pages readROSROS · Nemotron 3 Nano on Nebius Token Factory
Text · FindingTool calls right 84%, up from 71%3 backedClaude CodeClaude Code · Opus · from annotation_guide.pdf p.3 and eval_tools.py
Claude CodeWorking
Agentread the guide, p.3
writing eval_tools.py
Notebookfinetune_lora.ipynbCodexCodex · ran in the Python sandbox · epoch 2 of 3

Make the connections.

Put the plan next to the data. Group the evals. Link the chart to the claim it backs. The agents work from the connections you make.

backs
Notion pageTraining plan14 steps · read live
Dataros_sessions.jsonl12,400 turnsGemini CLI
Group · Evals · 2 blocks
Text · FindingTool calls right 84%, up from 71%
Chartval_loss.png

Direct the work.

Ask ROS from the bar on the canvas. It plans the work and hands each task to the right agent. Changes arrive as proposals: accept, change or undo. You set how much lands without asking.

ROS runs on NVIDIA Nemotron 3: hosted free on Nebius Token Factory, or a 3 GB download on your Mac.

NVIDIA Nemotron 3 Nebius Token Factory Runs offline too
Fine-tune Nano on our sessions and check its tool calls
ROS proposes
1Label 1,200 turns against the guideGemini CLI
2Train the LoRA, three epochsCodex
3Check tool calls and write the findingClaude Code
Unknown: no eval set for the new tools yetStart
3 changes to reviewClaude Code · 14:06
Add Finding: Tool calls right 84%, up from 71%Text
Add val_loss.pngChart
Draft report.md from the Evals groupText
Accept allShow on canvas
Ask every time Approve for me Full access

Yours, on your own machine.

Open source under MIT. It runs on your computer and talks only to the providers you choose.

  • Your keys stay with you.
  • Use the agent plans you already have. Claude Code, Codex and the others sign in as they do today.
  • Works offline. With local models and the local ROS.
  • Plain files. Markdown, notebooks and CSV that other apps open too.

Run your agents in the open.

Mac and Windows app Download for Mac

Apple silicon · Intel · Windows

Terminal 
curl -fsSL https://ros-ai.io/install.sh | sh

Mac, Linux or Windows under WSL · Node 24 · From source

Try the live demo