MIT Licensed · Self-hostable · Open source

AI agent sandboxes,
under your control.

Launch Claude Code, Codex, opencode, and Gemini CLI in fully managed cloud sandboxes. Full session visibility, cost control, and a live browser terminal.

  • Isolated sandbox per session
  • Your API keys, your models, your data
  • No servers to provision — boots in seconds
Works with:Claude CodeCodexopencodeGemini CLI

100% open source · MIT License · Self-hostable · Managed sandboxes

How it works

From zero to agent in minutes.

No servers to provision. No infrastructure to manage. Just a managed sandbox and your agent of choice.

01

Start a session

Sign in, pick a project, and choose your agent. No servers to provision, no infrastructure to manage — sandboxes are fully managed for you.

02

It boots in seconds

Each session spins up its own isolated cloud sandbox in seconds, with your repo pre-cloned and dependencies ready to go.

03

Watch your agent work

A live xterm.js terminal streams the agent output directly to your browser. Interrupt, redirect, or observe — full control, real time.

Live sessions

Watch your agent work in real time

Full PTY output streamed directly to your browser.

raziel — claude-code — 80x24

$ claude --add-dir . "Refactor auth middleware to use JWT RS256"

Initializing Claude Code session...

tool_use: read_file(path="middleware/auth.ts")

result: 142 lines read

tool_use: read_file(path="lib/jwt.ts")

result: 58 lines read

tool_use: edit_file(path="middleware/auth.ts")

result: replaced HS256 with RS256 verification, added key rotation

tool_use: edit_file(path="lib/jwt.ts")

result: exported generateKeyPair, signToken, verifyToken

Done. Auth middleware updated to RS256. 2 files changed.

$

"Finally, AI agents that don't need root on my laptop."

— A developer who values security

Features

Built different.

Real isolation. Real control. No managed fluff.

Isolated

Isolated per session

Every session runs in its own managed cloud sandbox — separate filesystem, network, and process space. Sandboxes can't reach each other or your host.

Any agent, any model

Claude Code, Codex, opencode, Gemini CLI — or just a shell. Your choice, your API keys.

No infra to manage

Sandboxes are fully managed and boot in seconds. Nothing to provision, patch, or keep running.

Live browser terminal

Full xterm.js PTY in the browser. Watch your agent work in real time, interrupt when needed.

Connect GitHub

Import repos as projects. Sessions are pre-cloned with your codebase ready to go.

Open source, MIT

Own your infrastructure. Self-host the whole stack. No vendor lock-in, ever. Star on GitHub →

Deploy in one click

Self-host in minutes.

Raziel is fully open source. Deploy the web app to Vercel or Fly.io — managed cloud sandboxes handle the agent runtime, with nothing to provision.

Requires: Clerk account · PostgreSQL (or Supabase)

Raziel — AI Agent Sandboxes