Self-hosted sandbox orchestrator.
Each sandbox is a Firecracker microVM with its own kernel, filesystem, and network. Idle VMs pause automatically and resume in microseconds. Runs on any Linux machine with KVM.
curl -fsSL bhatti.sh/install | sh Isolation without the overhead.
Real VMs. Not containers.
Every sandbox runs its own Linux kernel in a Firecracker microVM. Process isolation, filesystem isolation, network isolation. Not a namespace trick — a separate machine.
Memory snapshots. Not just filesystem.
When bhatti snapshots a VM, it captures everything: running processes, open file descriptors, TCP connections, in-memory state. Resume picks up exactly where it left off.
Three thermal states. Invisible to you.
Idle 30 seconds → warm (vCPUs paused, ~400µs resume). Idle 30 minutes → cold (snapshotted to disk, memory freed, ~50ms resume). Any API request transparently wakes it.
| Operation | p50 | p95 | p99 |
|---|---|---|---|
| Run a command | 1.22ms | 1.66ms | 1.79ms |
| Read a file (1KB) | 583µs | 757µs | 821µs |
| Write a file (1KB) | 1.06ms | 1.25ms | 1.48ms |
| Command on paused sandbox | 2.08ms | 5.85ms | 5.85ms |
| Command on cold sandbox | 40.6ms | 46.3ms | 46.3ms |
| 10 commands at once | 11.4ms | 18.6ms | 18.6ms |
| Create a sandbox | 1.44s | 1.45s | 1.45s |
| Diff snapshot | 32.0ms | 38.1ms | 38.1ms |
Measured on Hetzner AX102 (AMD Ryzen 9, x86_64, NVMe). Pi 5 numbers in the repo README.
How it works
Click any component to explore.
Built for
AI Coding Agents
Give LLMs a real, stateful environment to execute code, install packages, and run tests — with full VM isolation between tenants.
# REST API — what your agent calls
$ curl -X POST /sandboxes \
-d '{"name":"agent","cpus":2}'
# Streaming exec (NDJSON)
$ curl -N \
-H "Accept: application/x-ndjson" \
-d '{"cmd":["npm","test"]}' \
/sandboxes/$ID/exec
# Atomic file writes
$ curl -X PUT \
/sandboxes/$ID/files?path=/app.js \
--data-binary @app.js
# Truncated reads (offset/limit)
$ curl /sandboxes/$ID/files \
?path=/app.log\&limit=100Disposable Dev Environments
Spin up isolated Linux VMs with Node 22, git, and common tools pre-installed. Snapshot and resume across sessions.
# Shell with session persistence
$ bhatti create --name project \
--cpus 2 --memory 1024
$ bhatti shell project
lohar@project:~$ npm install
lohar@project:~$ npm run dev
# Ctrl+\ to detach — keeps running
# Reconnect later, scrollback replayed
$ bhatti shell projectMulti-Tenant Sandboxing
Each user gets isolated API keys, network segments, resource caps, and rate limits. VMs from different users cannot communicate.
# Per-user isolation at every layer
$ sudo bhatti user create --name alice \
--max-sandboxes 5
$ sudo bhatti user create --name bob \
--max-sandboxes 10 \
--max-cpus 4 --max-memory 4096
# Each user gets:
# - Dedicated bridge + /24 subnet
# - Scoped API key (SHA-256 auth)
# - Per-user rate limits
# - Encrypted secrets (age)Run it on your hardware.
Clone, install, serve. Any Linux machine with KVM. Builds both binaries from source, downloads a kernel, and creates an Ubuntu 24.04 rootfs. No Docker, no containers, no orchestration layer.
Requirements
- Linux (x86_64 or ARM64)
- Hardware virtualization (/dev/kvm)
- Root access for daemon (Firecracker requirement)
Tested on: Raspberry Pi 5, Hetzner AX-series, AWS Graviton metal.
# 1. Clone and install (builds from source, ~10 min first time)
$ git clone https://github.com/sahil-shubham/bhatti.git
$ cd bhatti && sudo ./scripts/install.sh
# 2. Start the daemon (requires root for kvm/network)
$ cd /var/lib/bhatti && sudo bhatti serve
[INFO] Starting bhatti daemon v0.1.0
[INFO] Listening on :8080
# 3. Create a user (API key shown once)
$ sudo bhatti user create --name alice
bht_live_abc123def456...
# 4. Install CLI on your laptop (macOS or Linux)
$ curl -fsSL bhatti.sh/install | sh
$ bhatti setup
API endpoint: https://your-server:8080
API key: ****
Testing connection... ✓ connectedDocumentation
Quickstart
CLI install, server setup, user management
→Architecture
System design, data flow, concurrency model
→Wire Protocol
Binary framing, connection lifecycle, auth
→Guest Agent
PID 1 init, PTY sessions, process management
→Thermal Management
Hot/warm/cold states, diff snapshots
→Networking
Per-user bridges, iptables isolation
→API Reference
REST/WebSocket endpoints
→CLI Reference
All commands and flags
→Testing
11K lines of tests, zero mocks for VM tests
→