What runs where
Your agents run with your privileges on your machine. This page says what Vendetta can do, what it can't, and what leaves the computer.
What Vendetta does and doesn't do
- Approvals are yours alone. A Claude Code session's permission dialog shows Allow once and Deny in the iPhone and web apps, and you answer it there or at the terminal. Telegram, Discord, agents, the Master, and schedules never answer one.
- Nothing uncertain is repeated. If a prompt, launch, or deletion may or may not have happened, Vendetta says so and leaves it. A failed schedule stays stopped until you check.
- Input goes to one exact process. Each is pinned and revalidated, so a stale button can't reach a replacement agent. Terminal output is treated as untrusted.
- Nothing listens on the network by default. The app port is on loopback until you publish it, every connection needs a paired device, and only the owner you configure can talk to the bots.
- History holds no conversations. Journals keep states and metadata, never prompts, messages, or tool arguments.
- It is not a sandbox. Agents on your computer, outside a container, run with your privileges. Agents you start by hand keep their own settings.
Approvals
Agents Vendetta starts on your computer ask before running commands by default. You can change the arguments they launch with, per provider, in the configuration. Launch arguments
Agents in containers don't ask. The container is the boundary. It holds a private clone of the repository, and your logins reach it through a proxy and are never copied in.
What leaves the computer
Nothing, until you turn something on. Each row is off until you set what the last column names.
| Connection | What goes out | Turned on by |
|---|---|---|
| Your agents' own traffic | Their providers, as before. Vendetta is not in that path. | Running the agent |
| Telegram or Discord bot | What those chats show. Bot chats are not end-to-end encrypted. | [plugins.telegram] or [plugins.discord] |
| Voice notes and dictation | The audio, to the transcription provider you pick. A local command sends nothing. | [plugins.speech], its transcription table |
| Replies read aloud | The reply's text, to the voice provider you pick. A local command or the device's own voice sends nothing. | [plugins.speech], its narration table |
| iPhone push | Alert text, to Apple's push service. | [plugins.app.apns] |
| Web push | Alert text, to the browser's push service. | Push notifications in the web client's Settings |
| ntfy | Alert text, to the topic you name. | [plugins.ntfy] url |
| GitHub | Requests to GitHub with your own token: your repositories, and git. | [plugins.github], then connecting in Settings |
| The app port | Nothing by itself. Paired devices connect in to it over the network you publish it on. | Publishing the port, for example on your tailnet |
| Webhooks | Nothing. Services you choose call in, with signed requests. | [plugins.webhooks] |
Pairing and revoking
Each phone and each browser is paired once. A pairing code works once, within 10 minutes, and every connection after that needs the device's token. Messages and media stay in memory on your machine, and each device keeps its own history.
$ vendetta app pair # a new code
$ vendetta app devices # list paired devices
$ vendetta app revoke ID # disconnect oneWhat a person you add can reach
- A person is a Vendetta of their own on your computer: their own worker, schedules, app endpoint, and paired devices, running under your account.
- Their agents run only in their container projects, in rootless Podman, and never directly on your computer. A container sees its own clone and its own sockets.
- Their Vendetta never sees your agents, and yours never sees theirs, apart from a project one of you shares.
- In a shared project they can prompt agents, answer questions, and start new ones. Stopping and deleting stay with the person who shared it.
- None of your skills or MCP servers reach them until you pick each one. A shared MCP server runs on your side, as you.
- You set each person's limits: memory, CPUs, and how many containers.
- Your own agents run as you, so they can read the folders of the people you add.
People and the security page in the docs have the details.