A terminal workspace for coding agents

Run agents in your terminal.
Answer from your phone.

Bring Claude Code, Codex, and your other agent CLIs into one workspace. Run them on your laptop or an SSH host. Pair your phone to answer supported requests.

Open source · Apache-2.0 · v0.22.0

A paired phone shows an agent's request to run bun test, with an Allow button
Narrated. Sample workspace with scripted agent output.
Read the preview

Run Claude, Codex, or any agent in the terminal you already use. When a supported agent needs permission, the request waits in one inbox. Away from your desk? Approve it from your phone. The agent continues in your terminal. Your agents keep working, wherever you are.

Start with one project

Keep your usual
agent commands.

  1. 1. Install the host CLI. Use macOS 13+ (Apple silicon or Intel) or Linux (x86_64 or ARM64).
  2. 2. Open your project. Run vibeke in its directory. Complete first-run setup.
  3. 3. Start an agent. Run claude, codex, or your usual command inside the pane.
Follow the quickstart

Install minisign first: brew install minisign on macOS, or use your Linux package manager.

install the CLI · requires minisign
curl -fsSL https://vibeke.dev/install.sh | sh
in your project directory
cd your-projectvibeke
Installation requirements and verification
01 / While the agents run

Find the agent
that needs an answer.

Claude edits a test. Codex reviews a diff. When an agent needs an answer, the inbox shows its request and task.

[01]

Know which agent needs you.

Keep agent questions and permission requests in one inbox. Answer through the supported agent integration.

Claude Code · Codex · pi / omp
Agents and interactions
[02]

Reconnect to work in progress.

Agents keep running after a client disconnect or server restart. A host restart stops the processes.

your work stays on the host
What survives a restart
[03]

Give each task its own branch.

Use a separate worktree for each task. Review its changes and check results before you accept the work.

separate files · reviewable changes
Tasks and review
See a useful workflow

From the terminal
to the next step.

Answer from your phone.

Pair your phone with a running host. Read an agent request, then allow or deny it. Your agent keeps working on the host.

Watch phone approvals

Requests depend on the agent integration. The host must stay awake and connected. Read the guide

Preview work on your devbox.

Run an app on an SSH host. Open its preview beside your agent in your local terminal.

Watch a remote preview

Browser panes need Chromium and a terminal that supports graphics. Read the guide

Example · agent mentions

Ask another agent for a review.

Name a pane @reviewer. Send it a request from another agent pane in the same workspace.

@reviewer please check the changes in src/auth.rs

Mentions work with Claude Code and Codex integrations. Messages wait until the receiving agent is idle.

Set up agent mentions
Example · page annotations

Show your agent what to change.

  1. Open your app in a browser pane.
  2. Choose Annotate page in the command palette.
  3. Pick an element. Write a note such as “Make this button easier to read.”
  4. Send the notes to your agent.

The notes include the selected element and a screenshot. They wait while the agent is busy.

Use page annotations
Ways to connect

One host.
Your choice of screen.

Start in your terminal. Add the desktop app or pair your browser with the running host when you want another view.

~/ One session. Three ways to connect.

Run agents in your terminal.

Screenshots with sample data.
TUI
Vibeke terminal with workspaces and agent output from a session recovery task.

Desktop app

Available for macOS, Linux, and Windows. The desktop app connects to a host. Local use also requires the CLI. Windows connects to a supported remote host.

Download desktop

Browser and phone

Open Devices on the host, then choose Pair a phone. Confirm the pairing to connect. Your host must stay awake and online.

~/ Product tour · 2:41

See it at work.

Watch agents work in the terminal on two machines. Preview a dev server. Hand work to another computer. Answer requests from the desktop app and a phone. The video uses sample data.

02 / From your shell

Start locally.
Attach over SSH.

Start a workspace.

Start Vibeke in your project. Complete setup, then run your usual agent command in a pane. Use the command palette for workspaces, splits, and the inbox.

Follow the guide
~/workspace — local
# In your project directorycd your-projectvibeke # Inside a Vibeke paneclaude  # or codex, pi, omp
Host mode uses your permissions. Use a supported sandbox, container, or VM to limit access.
04 / Documentation

Start with
one workspace.

Use the quickstart to start your first agent. The reference lists commands, API methods, and configuration settings.

Open the docs
Before you install

Common questions.

Which computers can run Vibeke?+

The host CLI runs on macOS 13 or later and Linux. macOS supports Apple silicon and Intel. Linux supports x86_64 and ARM64. The Windows desktop app connects to a supported remote host.

Can I use my existing agents and accounts?+

Yes. Run your usual agent commands in a pane. Agent and model charges remain with your providers. Structured requests and other features depend on the integration.

How do the desktop and browser apps connect?+

Both connect to a running Vibeke host. The desktop app does not bundle the CLI. Install the CLI for local use, or pair with a remote host. Pair your browser or phone from the host.

What survives a restart?+

Holder processes keep agents active after a server restart or client disconnect. Screen recovery can be incomplete. A host restart or holder failure stops the processes.

Does every task run in a sandbox?+

No. Host mode uses your permissions. Use a supported sandbox, container, or VM to limit access. Worktrees separate file changes. They do not prevent access.

Try Vibeke in your next coding session.

Start with one project and your usual agent.

Want to come back later? Save the repository with a star.