Organization repository on GitHub
Secure and persistent computer for AI agents
- Forks
- 84
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Demos & Experiments
Celesto
Sandboxed virtual machines for agents: each sandbox is a microVM that starts in about 500 ms, keeps files and state between sessions, and runs locally or in Celesto Cloud behind a CLI and Python SDK.
- Stars
- 1k
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Author
- CelestoAI
- Last verified
- Oct 2026
About this project
OpenMuse · Quickstart · Python API · Agents and automation · Runtimes · Examples · Docs · Discord. Celesto gives an AI agent its own computer for running code, browsing the web, and using desktop apps. You can run that computer on your machine during development or in Celesto Cloud for remote and production work.
Quoted from the project's README. Read the full README.
- Languages
- Python
Highlights
- Each sandbox is its own lightweight VM, so agent code runs behind a real boundary rather than as a process on your machine.
- Runtimes range from a bare command sandbox to Chromium with CDP, a Linux desktop, Windows, and a macOS preview.
- Browser sandboxes expose CDP, viewer, and display URLs for Playwright, live watching, and VNC or computer-use clients.
- Automation gets machine-readable output on every command, with exit codes, error fields, and a recovery command on failure.
Watch out
Setup is not frictionless: the local path needs virtualization support, with QEMU installed through Homebrew on macOS and a possible sudo prompt on Linux, and the one-line installer pipes a shell script from the network. Local telemetry is on by default, shows a notice, may infer approximate location from the request IP, and is turned off only by running a config command or setting an environment variable. Cloud sandboxes need an API key. Coverage is uneven: the macOS desktop is a preview with a CLI path but no Python API in the runtime table, host mounts are read-only or writable per directory, and outbound restriction is limited to allowing specific IPv4 ranges on Linux with Firecracker. The first Linux desktop start downloads and verifies an image whose size is not stated. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
The stated reason is the boundary. Because the agent runs in a separate virtual machine instead of a process on your computer, the README claims a stronger boundary for untrusted code, and it starts a ready microVM in about 500 ms without pulling an image on each start. If your threat model is a runaway script rather than genuinely hostile code, that cost may not pay for itself.
The local path installs virtualization software, not just a Python package: QEMU via Homebrew on macOS, and on Linux setup may ask for sudo. Cloud skips that, needing only the Python package and an API key, and exposes the same Computer API. So the deciding factor is whether you want a VM stack on your own machine or a key and a remote bill.
Both, through separate runtimes. A browser sandbox exposes a CDP endpoint for Playwright or another CDP client, plus URLs for live viewing and for screen control over VNC. A Linux computer sandbox gives a visible desktop with Chromium, a terminal, a file manager, and a text editor. Windows and macOS desktops exist too, macOS as a preview. The default computer object is commands, code, and files only.
Local CLI and Python SDK usage is described as sending limited installation and feature counts after displaying a notice, with PostHog potentially inferring approximate location from the request IP address. It can be disabled with a config command or an environment variable. Whether cloud sandboxes transmit anything comparable is not documented in the README.
Five. Minimal computer covers commands, code and files through Computer() and celesto computer. Browser is Chromium with CDP, screenshots and a live viewer through Celesto.browser(). Linux desktop is a full desktop with several GUI apps. Windows computer needs Celesto(os="windows", ...) plus --image PATH. macOS desktop is for app and installer tests on Apple Silicon but is a preview: the CLI path exists and there is no Python API entry for it in the runtime table. Use Celesto(...) directly when you need backend, communication channel, guest OS, mounts or network policy.
Every command takes --json and returns ok, command, exit_code, data and error, with a nonzero exit code on failure and a recovery command in the error when one can be offered. celesto doctor --strict turns warnings into a failed check, unique names matter for concurrent jobs, and exec is the command for automation while shell, ssh and desktop are reserved for interactive work. There are presets too: celesto codex start, plus Claude Code, Pi, Hermes, OpenCode and OpenClaw.
Files and state survive across sessions. Host directories can be mounted read-only or writable, per directory. Snapshots pause and restore memory, disk and active processes. Network policy can disable outbound access or allow specific IPv4 ranges, on Linux with Firecracker. Linux, Windows 11 and a macOS preview are supported, and the first Linux desktop start downloads and verifies its image while later starts reuse the cached copy, so Docker is not required.
More Jev resources like this
147GitHub stars
Apache-2.0 open-weight stand-in for a hosted decision model: Qwen3.5-4B with a distilled adapter scores noul, choice and score questions through a systemone-shaped server, with no generated tokens.
Jev Demos & Experiments#community#jevbest
Community
allebee/jevk5Python
allebee/jevk5Python56GitHub stars
macOS agent loop that moves a real USB HID mouse and keyboard under a vision model, with an experimental decision-model lane picking the next operation and screen target from an element table.
Jev Demos & Experiments#community#jevbest
Community
affirmitv/ghosthandsPython
affirmitv/ghosthandsPython49GitHub stars
StarCraft II research repo combining Jev action selection with optional Astra plans across full-game macro control and SMAC-Hard micromanagement, with videos, replays, and a paper.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
sc2musa/Jev_StarPython
sc2musa/Jev_StarPython22GitHub stars
Terminal harness that checks ROS 2 robots with deterministic health tests, capability descriptions, SQLite history, and optional Jev evidence routing.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
YueBit/robodiag-harnessPython
YueBit/robodiag-harnessPython19GitHub stars
MCP server where one browser_goal call hands a URL, a goal, and code-checked assertions to a server-side loop driven by a decision model over live page elements.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
jiawei686/jev-ultrafast-mcpPython
jiawei686/jev-ultrafast-mcpPython16GitHub stars
Local Qwen3.5 decision models in 0.8B and 2B sizes with fitted temperatures, systemone-shaped server, playground, guard, MCP tools, and eval harness.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
lawrence3699/jev-stylePython
lawrence3699/jev-stylePythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.