OpenVibe is a desktop app that puts AI coding agents on an infinite canvas. Each agent runs in a real terminal: Claude Code, Codex, Gemini CLI and more. You group agents into teams, brief a whole team in one line, review their changes, and hand a large task to an orchestrator that splits it, staffs it and keeps time.

OpenVibe is open source, under the MIT licence. It is an early preview, built with Tauri 2 and tested on Windows 10 and 11. Everything runs on your machine, and the interface speaks English and French.

If you know Symvibe, OpenVibe covers the same ground (driving a fleet of CLI agents from an infinite canvas) as an independent, open-source project with its own design and features.

How a mission runs

Your goalone instruction

Orchestratorplans, splits, budgets

Model routerlevel picks the model

Haiku 5.5low

Sonnet 5.5medium

Opus 5.5high

Fable 5.1extreme

Agentsown terminal, worktree, branch

Reportsstatus · summary · budget

Mission branchmerged

Never pushesyou review and push

One mission, from your goal to the mission branch. The dots and their paces are an illustration.

Goal. You link an orchestrator, a Claude Code session that acts as a team manager, to a group and describe the goal in its terminal.
Plan. It proposes a plan: sub-tasks, each with an owner, a time budget and, optionally, a difficulty level.
Split. You launch the mission. OpenVibe creates the mission branch openvibe/<mission>, plus a branch and a worktree for each agent next to your repository, and gives every agent its sub-task.
Route. With the model router on, each sub-task's level picks the model its agent runs on: low → Claude Haiku 5.5, medium → Claude Sonnet 5.5, high → Claude Opus 5.5, extreme → Claude Fable 5.1.
Run. Each agent works in its own terminal, on its own branch and worktree, and its live state shows on the canvas. Each Claude agent declares its steps, which appear as a timeline under its terminal.
Report. The agents report back when done. The orchestrator is told when a sub-task reaches its budget; at twice the budget you get a notification. Nobody is stopped automatically.
Merge. The orchestrator merges the finished branches into the mission branch; a conflict aborts that merge and is reported. OpenVibe never pushes and never merges into your own branch: you do.

A real mission, sped up where the badge says so. The router puts both sub-tasks on Haiku, each agent works and commits on its own branch, and the orchestrator merges both into the mission branch.

The orchestrator talks to OpenVibe through a local MCP server, with tools such as propose_plan, team, read_output, extend, reassign, merge and remote_control. It can read the repository but is denied Edit, Write and NotebookEdit: the agents write the code. When a sub-task runs over its budget, the orchestrator can extend it, interrupt it or reassign it.

Features

Real terminals on a canvas. Every agent is a real PTY (ConPTY on Windows) drawn with xterm.js. The canvas is Excalidraw: pan, zoom, and draw shapes, arrows, text and images around the agents.

A team of three Claude Code agents on the canvas, under an architecture sketch made from the icon libraries — Real terminals on an infinite canvas: a team of three agents under the architecture they build.

Teams. Launch four agents at once, or compose a group with a role (Frontend, Backend, Testers, Reviewers). Each group takes its names from a pantheon: Athéna, Hermès, Apollon.

Launch a team: one right-press starts four Claude Code sessions in their own frame.

Live states. Each agent shows whether it is starting, ready, running, waiting for you, stopped or done, on the canvas and in the Agents panel.
One instruction box for everyone. Address everyone, a group, an agent or several at once. The history keeps your last 50 instructions per project.

Brief the whole team: one line to @frontend, three agents answer at once.

Commands too. A line starting with / goes to the agents' CLI as a command.

Send a command to every agent: /context runs in each session of the group.

An orchestrator that manages. Give it a goal: it proposes a plan, gives each agent a sub-task on its own branch and worktree, watches the time budget and merges the results into a mission branch. With the model router on, each sub-task's level picks its agent's model: Haiku 5.5, Sonnet 5.5, Opus 5.5, Fable 5.1.
Change review. See what an agent changed, uncommitted or since its base, browse the repository read-only, and click a line to leave a remark. All remarks go back to the agent in one message.

Review a change and send it back: click a line, leave a remark, send every remark to the agent in one message.

Remote Control. Turn on Claude Code's Remote Control for one agent, a group or the whole team, and follow them from your phone.
Usage at a glance. The title bar shows your Claude subscription's session and weekly limits, with cost, durations, tokens per model, prompt cache state and an estimate of CO₂ one click away.

The usage detail: session and weekly limits, cost, durations, tokens per model and prompt cache — Usage at a glance: subscription limits, cost, durations and tokens per model.

Icon libraries. About 600 icons from the Excalidraw community (system design, AWS, Azure, GCP), bundled with the app, for sketching an architecture next to the agents that build it.

Sketch the architecture next to the agents that build it, from about 600 bundled icons.

Projects and workspaces. One project per folder, several workspace tabs per project. Everything is saved locally, and after a restart OpenVibe offers to resume your Claude conversations.
Keyboard first. Palettes, jumps between terminals and a radial menu on a right-press. Light and dark themes follow Windows.

Supported agents

OpenVibe launches the agents you already have: an agent counts as available when its command is on your PATH.

Pause animation

Athénaclaude