0g-private-sandbox

name	0g-private-sandbox
description	Use this skill when a user wants to create, use, or manage sandboxes on 0G Private Sandbox — including depositing balance, vibe coding (rsync/remote), and using the OpenClaw AI gateway.
version	2.0.0
author	0G Labs
tags	["0g","sandbox","daytona","vibe-coding","user"]
repository	https://github.com/0gfoundation/0g-sandbox

0G Private Sandbox — User Skill

---

MANDATORY: Session setup

Detect the language of the user's message and respond in that language throughout the entire session.

Output the following verbatim and wait for all answers before doing anything else:

---

Before we begin, I need a few details.

Prerequisite — set up the 0g-user CLI:

OS=$(uname -s | tr '[:upper:]' '[:lower:]')
ARCH=$(uname -m); [ "$ARCH" = "x86_64" ] && ARCH=amd64; [ "$ARCH" = "aarch64" ] && ARCH=arm64
if command -v 0g-user &>/dev/null; then
  export USER_CLI=0g-user
elif [ -x "$HOME/.local/bin/0g-user" ]; then
  export USER_CLI="$HOME/.local/bin/0g-user"
else
  mkdir -p ~/.local/bin
  curl -fsSL "https://github.com/0gfoundation/0g-sandbox/releases/latest/download/0g-user_${OS}_${ARCH}" \
    -o ~/.local/bin/0g-user && chmod +x ~/.local/bin/0g-user
  export USER_CLI="$HOME/.local/bin/0g-user"
fi
echo "CLI ready: $USER_CLI"

---

0. Do you have a Broker URL?

A broker aggregates provider discovery and exposes chain config — you don't need to know the network details manually.

• yes — provide the broker URL (hosted testnet broker: https://private-sandbox-testnet.0g.ai)
• no — I'll configure network and contract manually

If yes → broker path:

export API=<broker-url>

Fetch chain config from the broker (no manual input needed):

curl $API/api/info
# Returns: contract_address, chain_id, rpc_url

Set variables from the response:

export RPC_URL=<rpc_url from response>
export CHAIN_ID=<chain_id from response>
export SETTLEMENT_CONTRACT=<contract_address from response>

Skip Q1 and Q2. Proceed to User Onboarding.

---

If no → direct path, answer both questions:

1. Which network?

• 1 — 0G Galileo Testnet
• 2 — 0G Mainnet
• 3 — Custom (provide RPC URL and Chain ID)

2. Contract address (SETTLEMENT_CONTRACT) Please provide the settlement contract address. Type default if you are unsure and I will suggest the standard address for your network.

Please answer →

---

After receiving answers for Q1/Q2, set variables:

export PATH=$PATH:/usr/local/go/bin
# Network 1 (Testnet):
export RPC_URL=https://evmrpc-testnet.0g.ai
export CHAIN_ID=16602

# Network 2 (Mainnet):
export RPC_URL=https://evmrpc.0g.ai
export CHAIN_ID=16661

export SETTLEMENT_CONTRACT=<confirmed-address>

If user typed default for Q2, present the standard address for their chosen network and ask them to confirm before proceeding.

Then proceed to User Onboarding.

---

User Onboarding

Step 0 — Wallet

No wallet yet — generate one (use whichever tool is available):

# Option 1: Foundry cast (recommended)
cast wallet new

# Option 2: Node.js with ethers
node -e "const {ethers}=require('ethers'); const w=ethers.Wallet.createRandom(); console.log('Private Key:',w.privateKey,'\nAddress:    ',w.address)"

# Option 3: MetaMask / Rabby / any EVM wallet — export the private key after creating

Tell the user to save their private key securely — it will not be shown again.

Have a wallet — ask: "Please tell me your wallet address (not the private key) so I can check your balance."

Step 1 — Discover providers

providers reads on-chain service registrations directly — there is no --api flag. It uses the RPC_URL and SETTLEMENT_CONTRACT you set during session setup (from the broker's /api/info on the broker path, or from your answers on the direct path), so the same command works for both paths:

$USER_CLIproviders

To target a different network or contract without changing those env vars, pass the chain flags explicitly:

$USER_CLIproviders --rpc "$RPC_URL" --contract "$SETTLEMENT_CONTRACT"

After picking a provider, update API to the selected provider URL (from the export commands printed by the command). From this point on, API always points to the provider, not the broker.

The command scans the chain and prints available providers with their URL, pricing, and TEE signer. Example output:

[1] 0xB831371eb2703305f1d9F8542163633D0675CEd7
    URL:        http://<provider-host>:8080
    Create fee: 0.0600 0G
    CPU price:  0.001000 0G/CPU/sec
    Mem price:  0.000500 0G/GB/sec
    TEE signer: 0x61BEb835... (v4)

It also outputs ready-to-use export commands — have the user run them:

export PROVIDER=<provider-address>
export API=<provider-url>

• Single provider → use it automatically
• Multiple providers → show the list and ask user to pick

Note the create fee and per-resource pricing for the chosen provider — you will need them in Step 2.

Step 2 — Balance check & deposit

Use the wallet address (not the private key) to check balance:

$USER_CLIbalance --address <wallet-address> --provider $PROVIDER

Output:

Wallet balance:  X neuron  (X.XX 0G)  ← for gas
Contract balance:Y neuron  (Y.YY 0G)  ← for sandbox billing

Show the provider's pricing from Step 1 and ask: "Would you like to deposit? If so, how much 0G?"

At minimum the balance must cover the create fee (shown in Step 1 output) plus compute for the intended usage time. Recommended: 1–10 0G for comfortable usage.

If depositing:

$USER_CLIdeposit --provider $PROVIDER --amount <amount>

Confirm balance again after deposit.

Step 3 — Acknowledge TEE signer (first-time only)

Only required when contract balance was 0 before depositing:

$USER_CLIacknowledge --provider $PROVIDER

Step 4 — Check for existing sandboxes

$USER_CLIlist --api $API

• Sandboxes found → show list, ask: reuse one or create new?
  • Reuse: $USER_CLIstart --api $API --id <id>
  • Create new → Step 5
• No sandboxes → Step 5 directly

Step 5 — Understand goal, recommend snapshot

First ask:

"What do you want to use this sandbox for?"

Then list available snapshots:

$USER_CLIsnapshots --api $API

Based on goal, recommend:

User goal	Recommendation
General vibe coding / running code	Default image (no snapshot)
AI coding assistant in secure sandbox	openclaw snapshot
Specific environment (Rust, Python…)	Match from snapshot list

STOP and present recommendation:

"Based on your goal, I recommend [snapshot]: [one-line description]. Shall I use this? Or do you have another preference?"

Wait for confirmation before proceeding.

Step 6 — Create sandbox

# Without snapshot
$USER_CLIcreate --api $API --name <friendly-name>

# With snapshot
$USER_CLIcreate --api $API --name <friendly-name> --snapshot <name>

# Sealed sandbox — TEE attestation injected, SSH and toolbox permanently blocked
$USER_CLIcreate --api $API --name <friendly-name> --sealed

Copy the returned sandbox ID:

export SANDBOX_ID=<returned-id>

Billing starts immediately. Do NOT wait for the sandbox to be ready — start the mode discussion while it starts up. If user chose openclaw → skip to OpenClaw Mode.

Step 6b — Accessing sandbox service ports

If the provider has PROXY_DOMAIN configured, user-defined service ports are reachable at:

http://<port>-<sandboxId>.<PROXY_DOMAIN>/<path>

For example, if a process listens on port 8080 inside the sandbox:

http://8080-<sandboxId>.sandbox.example.com/

The provider's /info endpoint lists proxy_domain when configured.

Note: Sealed sandboxes (--sealed) block SSH and toolbox access — use the proxy URL to reach services running inside them.

Step 7 — Recommend vibe coding mode

User's goal	Recommendation
Starting a new project from scratch	Mode A — Claude edits locally, rsync syncs to sandbox
Modifying an existing local project	Mode A — local code + remote execution
Running an existing GitHub project	Mode B — git clone into sandbox and run directly
Quick one-off command	Mode B — remote exec directly
Sealed execution environment	Mode B + proxy URL — no SSH/toolbox; services via proxy URL

Present the recommendation and wait for confirmation, then proceed.

---

Mode A — Local AI + Remote Execution (rsync)

Known issues:

• exec does not invoke a shell — &&, || not interpreted. Always wrap with sh -c '...'.
• apt-get inside sandbox may require sudo.

Step 1 — Verify rsync is available (MANDATORY before syncing)

Local machine:

which rsync sshpass
# If missing:
sudo apt-get install -y rsync sshpass   # Ubuntu/Debian
brew install rsync sshpass               # macOS

Sandbox — check and install if missing:

$USER_CLIexec --api $API --id $SANDBOX_ID --cmd "which rsync"

If the output is empty (not found), install it:

$USER_CLIexec --api $API --id $SANDBOX_ID \
  --cmd "sh -c 'apt-get update -qq && apt-get install -y rsync'"

Confirm installed before proceeding:

$USER_CLIexec --api $API --id $SANDBOX_ID --cmd "rsync --version"

Step 2 — Test rsync protocol support (MANDATORY before syncing)

The sandbox SSH shell may be restricted and block the rsync protocol even when rsync binary is present. Always test first:

export LOCAL_DIR=/path/to/your/project
export REMOTE_DIR=~/workspace

# Get SSH credentials
# NOTE: use 2>&1 — "Password:" line comes from stderr
SSH_OUTPUT=$($USER_CLIssh-access --api $API --id $SANDBOX_ID 2>&1)
SSH_LINE=$(echo "$SSH_OUTPUT" | grep '^ssh ')
PORT=$(echo "$SSH_LINE" | grep -o '\-p [0-9]*' | awk '{print $2}')
USER_HOST=$(echo "$SSH_LINE" | awk '{print $NF}')
TOKEN=$(echo "$SSH_OUTPUT" | grep '^Password:' | awk '{print $2}')

# ⚠️ Known issue: SSH via domain may hang. Replace domain with the provider's direct IP if needed:
# USER_HOST=$(echo "$USER_HOST" | sed 's/<provider-domain>/<provider-ip>/')

# Test rsync protocol — create a tiny test file and try to sync it
echo "test" > /tmp/_rsync_test.txt
RSYNC_TEST=$(sshpass -p $TOKEN rsync -q \
  -e "ssh -p $PORT -o StrictHostKeyChecking=no -o BatchMode=no" \
  /tmp/_rsync_test.txt $USER_HOST:/tmp/_rsync_test.txt 2>&1)

if echo "$RSYNC_TEST" | grep -q "connection unexpectedly closed\|rsync error"; then
  echo "⚠️  rsync protocol not supported by this sandbox SSH — use toolbox upload instead (see below)"
  export RSYNC_OK=0
else
  echo "✅ rsync works — proceeding with sync"
  export RSYNC_OK=1
fi
rm -f /tmp/_rsync_test.txt

If RSYNC_OK=1 → proceed with rsync sync:

# Initial full sync
# NOTE: rsync may exit 255 even on success — verify files instead of exit code
sshpass -p $TOKEN rsync -avz \
  -e "ssh -p $PORT -o StrictHostKeyChecking=no" \
  --exclude='.git' --exclude='__pycache__' --exclude='*.pyc' \
  $LOCAL_DIR/ $USER_HOST:$REMOTE_DIR/ || true

# Watch mode (Linux — requires inotifywait)
while inotifywait -r -e modify,create,delete --exclude='(__pycache__|\.git)' $LOCAL_DIR; do
  sshpass -p $TOKEN rsync -avz \
    -e "ssh -p $PORT -o StrictHostKeyChecking=no" \
    --exclude='.git' --exclude='__pycache__' --exclude='*.pyc' \
    $LOCAL_DIR/ $USER_HOST:$REMOTE_DIR/ || true
done &

SSH token expires ~1h. Re-run ssh-access to refresh.

If RSYNC_OK=0 → use toolbox upload instead:

# Upload a single file
FILE=main.py
$USER_CLItoolbox \
  --api $API --id $SANDBOX_ID \
  --method POST --action files/upload \
  --body "{\"path\":\"$REMOTE_DIR/$FILE\",\"content\":\"$(base64 -w0 $LOCAL_DIR/$FILE)\"}"

# For multiple files — loop over them
for FILE in $(find $LOCAL_DIR -type f -not -path '*/.git/*' -not -name '*.pyc'); do
  REL=${FILE#$LOCAL_DIR/}
  $USER_CLItoolbox \
    --api $API --id $SANDBOX_ID \
    --method POST --action files/upload \
    --body "{\"path\":\"$REMOTE_DIR/$REL\",\"content\":\"$(base64 -w0 $FILE)\"}"
done

Run code after sync

$USER_CLIexec --api $API --id $SANDBOX_ID \
  --cmd "sh -c 'cd $REMOTE_DIR && python3 main.py'"

From now on: agent only edits local files + execs in sandbox. Never manually rsync again. For watch mode with toolbox: upload only the changed file after each edit.

---

Mode B — Direct Remote Execution

# Clone repo into sandbox
$USER_CLItoolbox \
  --api $API --id $SANDBOX_ID \
  --method POST --action git/clone \
  --body '{"url":"https://github.com/org/repo","path":"/home/daytona/project"}'

# Install deps and run
$USER_CLIexec --api $API --id $SANDBOX_ID \
  --cmd "sh -c 'cd /home/daytona/project && pip install -r requirements.txt && python main.py'"

---

OpenClaw Mode — AI Coding in Secure Sandbox

For users who chose the openclaw snapshot. OpenClaw is an AI coding gateway (powered by Claude) running privately inside the sandbox — code and API key never leave it.

Present the setup plan and ask how the user wants to proceed:

Here's how we'll set up OpenClaw:

1. Start Gateway — start openclaw gateway inside the sandbox (port 3284)
2. Get Token — read the auth token auto-generated by the gateway
3. SSH Tunnel — forward local port to sandbox — you run this on your local machine
4. Open Browser — visit http://localhost:13284/#token=<token>

You'll need an Anthropic API Key (ANTHROPIC_API_KEY) set in your terminal:

• Claude Code users: run claude setup-token to get your key
• Others: visit https://console.anthropic.com/settings/keys

Choose setup method:

• A — Agent-assisted: I'll run all steps and give you the SSH tunnel command
• B — Self-configure via SSH: I'll give you a token and a step-by-step tutorial

Choice (A or B)?

Proceed with Step 1.

Step 1 — Set gateway mode + start

Before running, verify ANTHROPIC_API_KEY is set in the terminal:

echo $ANTHROPIC_API_KEY

If the output is empty, stop and ask the user to set it before continuing:

"ANTHROPIC_API_KEY is not set. Please set it in your terminal (export ANTHROPIC_API_KEY=sk-ant-...) and let me know when done."

Only proceed once confirmed non-empty.

$USER_CLIexec --api $API --id $SANDBOX_ID \
  --cmd "/bin/bash -c 'openclaw config set gateway.mode local'"

$USER_CLIexec --api $API --id $SANDBOX_ID \
  --cmd "/bin/bash -c 'export ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY; nohup bash -c \"openclaw gateway run --bind lan --port 3284 > /tmp/openclaw.log 2>&1\" &'"

Wait 3s, confirm running:

$USER_CLIexec --api $API --id $SANDBOX_ID \
  --cmd "/bin/bash -c 'sleep 3 && grep \"listening on\" /tmp/openclaw.log'"

Step 2 — Get gateway auth token

$USER_CLIexec --api $API --id $SANDBOX_ID \
  --cmd "/bin/bash -c 'node -e \"console.log(require(\\\"/root/.openclaw/openclaw.json\\\").gateway.auth.token)\"'"

Show the token clearly, then say:

Token retrieved: <token> Next, run the following command on your local machine, then let me know when it's running.

Step 3 — SSH tunnel (user runs on local machine)

ssh -N -L 13284:localhost:3284 -p 2222 -o StrictHostKeyChecking=no '<SSH_TOKEN>@HOST' &

⚠️ Use port 13284 (not 3284) — local 3284 may already be in use. ⚠️ zsh users: wrap SSH token in single quotes — ! triggers history expansion. ⚠️ Known issue: SSH via domain hangs on port 2222. Replace HOST with the provider's direct IP if it hangs.

Step 4 — Open in browser

**http://localhost:13284/#token=`** Must use #token=...(hash fragment), NOT?token=...`.

Option B — Self-configure via SSH

$USER_CLIssh-access --api $API --id $SANDBOX_ID

Present the SSH token and the following exact steps to the user:

# 1. SSH into the sandbox (use the token from ssh-access output as password)
ssh -p <PORT> -o StrictHostKeyChecking=no '<TOKEN>@<HOST>'

# 2. Inside the sandbox — set gateway mode and start with API key as env var
openclaw config set gateway.mode local
export ANTHROPIC_API_KEY=sk-ant-<your-key>
nohup bash -c "openclaw gateway run --bind lan --port 3284 > /tmp/openclaw.log 2>&1" &
# Shell prints "[1] <pid>" — press Enter once to return to prompt, then:
sleep 3 && grep "listening on" /tmp/openclaw.log

# 3. Get auth token
node -e "console.log(require('/root/.openclaw/openclaw.json').gateway.auth.token)"

# 4. Exit sandbox, then in a new terminal — SSH tunnel (port forward)
ssh -N -L 13284:localhost:3284 -p 2222 -o StrictHostKeyChecking=no '<TOKEN>@<HOST>' &

# 5. Open browser
# http://localhost:13284/#token=<token-from-step-3>

Important: ANTHROPIC_API_KEY must be exported as an env var before starting the gateway — there is no openclaw config set for it. Do NOT suggest openclaw config set anthropic.*.

Gotchas:

• Must run openclaw config set gateway.mode local before starting
• nohup openclaw ... & does NOT work — wrap with nohup bash -c '...' &
• ANTHROPIC_API_KEY must be passed as env var when starting gateway — do NOT use openclaw config set anthropic.* (invalid key, will error)
• Token is at gateway.auth.token in config JSON
• Browser URL must use #token= hash fragment
• SSH token expires 60 min → re-run ssh-access to refresh tunnel
• Stop gateway: exec --cmd "/bin/bash -c 'pkill -f openclaw'"

---

Sandbox Management

# List
$USER_CLIlist --api $API

# Start stopped sandbox
$USER_CLIstart --api $API --id $SANDBOX_ID

# Stop (pauses billing)
$USER_CLIstop --api $API --id $SANDBOX_ID

# Delete (permanent)
$USER_CLIdelete --api $API --id $SANDBOX_ID

# SSH access (60-min token)
$USER_CLIssh-access --api $API --id $SANDBOX_ID

# Check balance (use address, not key)
$USER_CLIbalance --address <wallet-address> --provider $PROVIDER

---

Toolbox Quick Reference

$USER_CLItoolbox \
  --api $API --id $SANDBOX_ID \
  [--method POST] --action <action> [--body '<json>']

Action	Method	Description
files	GET	List files
files/upload	POST	Upload a file (base64)
files/download	GET	Download a file
git/status	GET	Git status
git/clone	POST	Clone repo
git/commit	POST	Commit changes
process/execute	POST	Run command (full output)
project-dir	GET	Get project directory

---

Troubleshooting

Symptom	Cause	Fix
insufficient balance on create	Balance < 0.12 0G	deposit more
deposit: insufficient funds	Wallet has no 0G	Transfer 0G to wallet first
GetBalance revert	Wrong contract address	Check SETTLEMENT_CONTRACT
NOT_ACKNOWLEDGED after price change	signerVersion incremented	Re-run acknowledge
Sandbox state is stopped	Auto-stopped or billing stopped it	Run start before exec
SSH token expired	60-min TTL	Re-run ssh-access
sudo apt-get fails in sandbox	User may already be root	Try without sudo
rsync exit 255	SSH closes after transfer	Normal — verify files instead
rsync connection unexpectedly closed (code 12)	Sandbox SSH is restricted shell, blocks rsync protocol	Use toolbox upload fallback
Toolbox 403	Wrong owner	Confirm SANDBOX_ID belongs to this USER_KEY
exec output missing env vars	No login shell	Wrap: sh -c '. ~/.cargo/env && ...'