VESPR · smart development environment (SDE)
The smart development environment.
Your team and your AI agents code in the same file. The room stays in charge.
Open source (AGPL-3.0). The repository link appears here when it is public.
Same file.
People and agents edit one file at once. Each has a name and a cursor you can see.
Ask the room.
Before an agent edits or runs something, the room gets one card. The first answer decides.
Know who wrote it.
Hover a character to see who wrote it, which model, and the request behind it.
what an SDE is
Not a chat box beside an editor. A room.
A smart development environment (SDE) is a room where people and AI agents work in the same files at the same time. Everyone has a name and a cursor. Agents ask before they act. Every edit keeps who made it and why.
What happens in a room
01 · same file
Everyone in the same file, at the same time.
VESPR is built on Yjs, a library for shared documents (a CRDT). Edits from people and agents merge without a lock and without a pull request in the way. Each person and each agent has a name and a colored cursor.
How a room works02 · ask the room
The room answers, not one person’s terminal.
When an agent wants to edit a file or run a command, it posts one card in the shared chat that says what it will do. Anyone in the room can answer. The first answer decides, and History keeps who gave it.
Modes and approvals03 · who wrote it
Every character remembers who wrote it.
Authorship is stored with the text: who wrote it, whether it was an agent, which model, when, and the chat message behind it. Hover a character to see it. It is a record the room keeps, not a signature, and it is not tamper-proof.
Tour of the appThese pictures are captures of VESPR on a throwaway demo stack: the agent’s replies there are scripted and no model key is used. Also built: History of every decision, the Code Map, session replay, Run for Python, skills, opening a GitHub repository. See the tour.
modes
You decide how far an agent may go.
Five modes. The mode belongs to the whole room, anyone in it can change it, and the change is written in the chat. Pick one and see what an agent may do on its own.
| Read a filecat, open, search | |
|---|---|
| Edit a filechange cart.py | |
| Run a safe commandpython3 -m unittest -q | |
| Run a risky commandnpm install, anything with a pipe | |
| Run a dangerous commandsudo, rm -rf on the home folder |
Preview: this models the table in the docs. It is not the app.
python3 -m unittest -q
A card nobody answers expires after five minutes and counts as no. Anyone with a login to the server can answer today; owner, editor and viewer roles are Planned.
measured · docs/MEASURED.md
Measured on a real model.
One task, three runs: fix the failing test in the sample project and run the tests again, in Auto mode, so no approval card was needed. The model was NVIDIA Nemotron 3 Super through Nebius Token Factory; the server and the agent ran on a Raspberry Pi.
| What | Range | Run 1 | Run 2 | Run 3 |
|---|---|---|---|---|
| First sign of workThe agent’s first step after “fix the bug”. | 0.5–0.9 s | 0.8 s | 0.9 s | 0.5 s |
| Edit in the shared fileThe fix appears for everyone in the file. | 1.5–2.9 s | 2.9 s | 1.8 s | 1.5 s |
| Tests greenThe fix, checked by running the tests. | 2.0–3.7 s | 3.7 s | 2.5 s | 2.0 s |
| Inline suggestionGhost text in the editor, median of five. | 623–678 ms | 623 ms | 678 ms | 652 ms |
first stepedit in the filetests green
Your numbers will differ with the model, the network and the task. Three runs of one task on one sample project is a measurement, not a benchmark. The method and every number are in docs/MEASURED.md in the repository.
what it runs on
A browser, a model, and your own computer if you want it.
Models
NVIDIA Nemotron 3 Super handles everyday chat, edits and inline suggestions. Nemotron 3 Ultra handles plan mode. Both run through Nebius Token Factory, and the model that wrote something is recorded with it.
nvidia/nemotron-3-super-120b-a12b
Your computer, if you choose
Commands run in a sandbox or on a runner on your own computer or a teammate’s. The runner connects out to the room, so you do not open a port. Runners are partly built: see the docs for what works.
About runnersA browser
The SDE opens in a browser. It has been checked in Chromium only; other browsers are not tested yet. A VS Code extension can join the same room if you prefer your own editor.
Run it yourselfopen source, and what is next
Open today. Honest about what is not built.
VESPR is open source under the AGPL-3.0 license. A model key is the operator’s own. The repository link will be here when it is public.
Run it on your computerOpen the live demo?
A room with a sample project and one agent. It runs in your browser; there is nothing to install.