Remote control for the agents on your computer

Vendetta finds every Claude Code, Codex, and OpenCode session on your machine and gives each one a chat on your phone. Answer its question, approve a command, or start new work from wherever you are.

Free and open source. Linux and macOS. iPhone app on TestFlight.

Who's it for

Eight agents across four projects. See which one is waiting and answer it from the train. See how it goes

Needs you

The agent that needs you is in a pane you aren't looking at

You start five agents and go make coffee. One finishes. One asks whether it may drop a table, and waits. Forty minutes later you find it.

Vendetta sorts every agent by what it needs from you. The one that is waiting is the first thing you see, on whichever screen is in front of you.

Questions and approvals

Its question arrives as buttons

When Claude Code, omp, or OpenCode asks you something, the question shows on your phone as a card with its options. Tap one or type your own answer. It goes to that exact process.

Claude Code's permission prompts arrive the same way. You see the command and why Claude wants to run it, then choose Allow once or Deny.

One conversation

Start at your desk and carry on from your phone

The agent runs on your computer the whole time, and the terminal, the browser, and the phone all show that one conversation. A prompt you type in the terminal appears in the chat. An answer you tap on the train is in the terminal when you sit back down.

Open whichever screen is nearest and keep going.

Voice

Talk to it, and have it read the answer back

Hold the mic and say what you want. The agent gets your words as a prompt, and the transcript sits under your voice note. Tap Listen on any reply to hear it read aloud, with code and tables skipped.

Voice mode makes it a conversation: it sends when you pause, reads the reply, and listens again. Start it with the Master and ask what's blocked while you drive.

Voice mode with the Master on iPhone: a sheet over the chat list with a microphone, Speak now, and Send Now

Marks

Mark the frame that's wrong

When an agent sends a video, scrub to the frame, draw a box on it, and say what's off. Mark an In and an Out to name a stretch. Open the narration and change a line to your exact words. The agent gets every mark numbered, fixes them, and sends the next cut.

Photos, audio files, and diffs take marks the same way.

Edit a video with an agent

Working

See the step it's on

Above the message box, a line shows how long the agent has been working, the command it is running, and where it is in its own list: "3 of 7, writing the migration". Open it for the whole plan.

Your lock screen keeps count of the agents that need you and shows the question that has waited longest.

Running

Your agents stay where they are

Your agents run in your own terminal, in tmux or outside it, with their own logins and settings. Vendetta finds them in the process list and reads what they report. It never wraps them or replaces their interface.

A Claude Code agent in its own terminal pane, partway through a turn: the prompt, its tool calls and their results, a test loop in progress, and the prompt box

Six places your agents show up

  • The overview in a terminal: agents as cards in four lanes, Needs you, Working, Ready and Ended today, with the next scheduled runs and today's tasks along the bottom
    Terminal
  • Vendetta's web client in a browser: the list of agents by project on the left, a running agent's conversation with its tool calls in the middle, and its plan, model, and context on the right
    Browser
  • iPhone
  • The desktop HUD: a card listing eight agents with their states, a to-do card with four tasks, and a notifications card where one agent needs approval and another asks a question
    Desktop
Terminal
A key you bind opens the chooser. One more jumps to the agent.
Browser
Every agent is a chat, with the overview beside it.
iPhone
The same chats, with voice notes and the share sheet.
Lock screen
How many agents need you, and the question that has waited longest.
Telegram and Discord
A topic or channel per conversation, with questions as buttons.
Desktop
Cards over your windows that never take the keyboard.

What else it does

All features

Pages in the chat

An agent can answer with a checklist, a comparison, or a form that opens in the chat and sends your choices back.

Read the diff

Committed work shows as a change. Mark a line and the agent gets your comment.

Prompts on a schedule

Every night, once on Friday, or when a webhook calls. Each run is recorded, and a failed one stops until you look.

Pick up where it ended

Ended conversations stay listed. Enter, or a message in the chat, restores the exact one.

A chat with an agent on iPhone: a chime alert saying the snapshot is restored, and a running 25-minute countdown to look at the upgrade check, with a cancel button

Alerts and timers

An agent on your computer can ring your phone in 20 minutes or count down a break. You hear it even with notifications muted.

Agents in containers

Run a project's agents in a container on a private clone of the repository.

Share from any app
Send a photo, a file, or a link from the iPhone share sheet into a chat.
Dictate at your desk
Press a shortcut and talk. The words go to the agent in the window you're in.

One agent to run the others

The Master is an agent you trust to list the others, read their replies, give them work, and answer their questions. Ask it what is blocked while you walk the dog.

A manager does the same for one project. Its team sits under it, and it tells you what you need to know. Approvals never go to either. Those stay with you.

  • Chief of staffmanager
    • Engineering lead3 agents
      • API
      • Web
      • Tests
    • Growth lead2 agents
      • Pricing page
      • Newsletter
    • Support lead1 agent
      • Inbox triage

Most of Vendetta is plugins, and you can write your own

Telegram and Discord, the desktop HUD, dictation, tmux and Kitty, webhooks, and the overview are each a separate program that Vendetta starts and keeps running. A plugin of yours uses the same socket they do, in any language.

It can add a page or a button to the apps, serve a command your agents call, and react when an agent asks, finishes, or fails. The example in the docs is one short Python file.

Write a plugin

notes.py
call("hello", {"plugin": "notes", "protocol": 1})

call("ui_set_contributions", {"contributions": [{
    "id": "note", "slot": "chat_menu",
    "label": "Add note", "on": {"action": "note"}}]})

call("ui_set_page", {
    "page": "notes", "title": "Notes", "root": "top",
    "nodes": {
        "top": {"kind": "list", "children": ["n1"]},
        "n1": {"kind": "row", "title": "Ship it"}}})

Share a project and work in it together

Share a project with someone and you both prompt its agents, answer their questions, and start new ones, each from your own phone. Every message says who sent it, and the agent answers that person. Stopping and deleting stay with you, and you can end the share whenever you like.

The people you add get a Vendetta of their own on your computer, with their own chats and their own Master. You don't see their chats and they don't see yours, apart from what you share.

Their app starts in Simple mode: their chats, by project, and nothing about models or providers. One switch in Settings shows everything.

It runs on your computer

Nothing listens on the network until you publish it, and then only paired devices connect. The history Vendetta keeps holds states and times. It never stores your prompts, the agents' messages, or tool arguments.

Data leaves your machine only through what you turn on, such as a Telegram bot or a transcription provider. What goes where

On your computer

Your agentsClaude Code, Codex, omp, OpenCode, Gemini CLI, Aider. Found in the process list, reporting through their own hooks, extensions, and Mods.
Worker, one per userAgents, history, schedules, project colours
Chooser and overviewIn your terminal
PluginsMessaging, tmux, Kitty, HUD, dictation
MasterAn agent you trust

Leaves it only when you turn it on

  • Your paired phone and browser
  • A Telegram or Discord bot
  • A transcription provider
  • Push notifications and ntfy

Works with the agents you already run

ProviderLive state fromReplies in chatQuestions as buttonsStart and resume
Claude Code 2.1.296+Native ModFrom its transcriptYesYes
CodexHooks, or its rollout filesFrom its transcriptNoYes
ompExtensionFrom its transcriptYesYes
OpenCodePluginTerminal snapshotYesYes
Gemini CLIHooksTerminal snapshotNoListed when you start it
Aider 0.86.2The vendetta-aider launcherTerminal snapshotNoListed when you start it

On macOS, omp and OpenCode agents are listed but report no state yet.

Free to run yourself, with a hosted version planned

Vendetta is open source under AGPL-3.0, and you run it on your own computer. The hosted version is for people who would rather not keep one on.

One email, when it opens.

Install

You need Linux or macOS and a Rust toolchain. Then pair your phone: vendetta app pair.

The full guide

$ git clone https://github.com/vendetta-sh/vendetta.git
$ cd vendetta && bash install.sh
$ vendetta notify start
$ vendetta tui

Questions

Do I need tmux?

No. Agents in any terminal are listed and can be messaged, stopped, and scheduled, and Claude Code, omp, and OpenCode take prompts natively. tmux adds pane navigation, live terminal previews, and managed agents, and Codex, Gemini, and Aider need it for remote prompts. Without a host

What leaves my machine?

Nothing, until you turn something on. Telegram and Discord see what those chats show, a transcription provider gets your voice notes, and ntfy and push get alert text. The apps connect to your own machine.

Can I approve a tool call from my phone?

For Claude Code, yes, in the iPhone app or the web client. The chat shows a card with the command or edit and why Claude asks, with Allow once and Deny. Other providers, and a subagent's dialog, you approve at the terminal, which the chooser can reach over SSH.

Is there an iPhone app?

Yes, on TestFlight. The web client works on any phone.

What happens if I message an agent while it's working?

It gets your message as a normal prompt, the same as typing it. Vendetta doesn't interrupt the turn, and a refused message is never replayed later.

Does a Master have to run?

No. It is optional, and nothing else depends on it.

An agent shows as unknown. Why?

Its integration isn't loaded. Install it, and restart the agent or trust its hooks. Provider integrations