Agents that get their own computer.
On yours.

Every conversation runs in its own Docker container on your machine, with a filesystem, a terminal and a browser preview, on the model you already pay for. Open source. One command.

curl -fsSL https://raw.githubusercontent.com/vonzio/vonzio/main/install.sh | bash

Steal the code (AGPL-3.0) macOS & Linux · Docker · 5 min

vonzio workspace: the agent's tool calls on the left, the Flask weather app it just built running in the Preview pane on the right — Unedited. Self-hosted v0.5.13, kimi-k2.7-code on Ollama Cloud, one prompt. The dashboard follows your system theme, so does this page.

Bring your own model.

vonzio doesn't resell tokens. Plug in the credentials you already have and keep your plan. Pick a different provider per agent, switch models per conversation.

AnthropicAPI key
Claude Pro / Maxclaude setup-token
ChatGPT Plus / ProCodex sign-in
Ollama CloudAPI key
OpenAI-compatibleOpenRouter · Azure · vLLM · LM Studio

Keys encrypted at rest, on your disk.
No quota caps, no markup.
Built on the Claude Agent SDK, so every provider gets the same tool loop.

The Preview pane showing the weather app the agent built, served from inside its container on port 8000 — The Preview pane, cropped from the capture above.

Each chat is a machine, not a textbox.

A fresh container with a real /workspace. The agent installs things, writes files, starts servers. You watch it in the Files pane, open a terminal next to it, and see whatever it's serving in the Preview pane. Idle chats pause to near-zero CPU and resume on the next message.

Files, terminal and browser preview on every conversation
Every tool call, file write and shell command stays in the thread
It asks you a question when it needs a decision, and waits
Voice dictation if typing the brief is slower than saying it

It doesn't wait for you to open a tab.

Talk to the same agent from Slack or Telegram. Put it on a schedule or behind a webhook with a playbook, give it a budget and a success criterion, and get the result back on Slack, Telegram, email or a webhook. Every run is a workspace you can open afterwards.

Slack and Telegram ship as plugins, on the same SDK you can use
Six built-in MCP servers: memory, notify, mail, calendar, local files, platform
Six agent templates to start from, or write your own system prompt in markdown

playbook · nightly-triage

trigger   cron 0 6 * * 1-5
agent     devops-helper · glm-5.3-flash
brief     read overnight CI failures, open a draft PR per fix
budget    $3.00 · 40 turns
done when every failing job has a PR or a reason
notify    slack #eng · summary + links

Open source

If you'd do it differently, fork it.

~ · what the installer actually prints

$ curl -fsSL …/vonzio/main/install.sh | bash
→ Installing: vonzio core v0.5.13 (latest release)
✓ docker 29.6.2 + compose v2
✓ .env created (random ENCRYPTION_KEY + BETTER_AUTH_SECRET)
→ Pulling prebuilt images (vonzio 0.5.13)…
✓ vonzio is up.
    Dashboard   http://localhost:3000   ← open this

Change the agent. The system prompt is a markdown file. Tools are JavaScript files in a folder.
Add a surface. Plugins declare the capabilities they need; routes, MCP servers, scheduled jobs, dashboard panels.
Run it for your team. AGPL-3.0: use it, change it, ship it. Host a modified copy for others and you share the changes back.

AGPL-3.0 · Node 22 · Postgres 17

Fork on GitHub Plugin guide →

Security, honestly

Containers, not microVMs.

Agents share the host kernel. You get encrypted credentials, auth on every route, a socket proxy in front of Docker, and an opt-in egress proxy that allowlists where containers can talk. You do not get kernel-level isolation from hostile code. Right for one developer, a trusted team, or a lab. For untrusted code, read the hardening guide and the threat model first.

Your agents deserve a real computer.

Install it, add a key, pick a model, hand something off. Five minutes, on your machine.