openclaw-aws-deploy（OpenClaw AWS 部署）

OpenClaw AWS Deploy Skill

Quick Start (Minimal Deployment ~$30/mo)

Prerequisites

• AWS credentials — any of these methods:

- --profile <name> flag (named AWS CLI profile) - .env.aws file in workspace root or skill directory (optional):

    AWS_ACCESS_KEY_ID=...
    AWS_SECRET_ACCESS_KEY=...
    AWS_DEFAULT_REGION=us-east-1

- Existing environment variables, AWS SSO session, or IAM role

• .env.starfish in workspace root (recommended) or skill directory:

  TELEGRAM_BOT_TOKEN=...     # from @BotFather (required)
  TELEGRAM_USER_ID=...       # your Telegram user ID (optional, enables auto-approve pairing)
  GEMINI_API_KEY=...         # from aistudio.google.com (optional, for Gemini models)

• aws CLI installed OR Docker for sandboxed access
• jq, openssl available

One-Shot Deploy

# From the skill directory:
./scripts/deploy_minimal.sh --name starfish --region us-east-1 \
  --env-dir /path/to/workspace

# Or with cleanup of previous deployment first:
./scripts/deploy_minimal.sh --name starfish --region us-east-1 \
  --env-dir /path/to/workspace --cleanup-first

This single command:

1. Creates VPC + subnet + IGW + route table
2. Creates security group (NO inbound ports — SSM only)
3. Creates IAM role with minimal permissions (SSM + Parameter Store + Bedrock)
4. Stores secrets in SSM Parameter Store (fetched at each service start — rewritten on each start, never stored in repo or static images)
5. Launches t4g.medium ARM64 instance with user-data bootstrap
6. User-data installs Node.js 22 + OpenClaw + configures everything
7. Runs smoke test via SSM
8. Saves all resource IDs to deploy-output.json

After Deploy

1. Message the Telegram bot — you'll get a pairing code
2. Approve pairing via SSM:

   aws ssm start-session --target <INSTANCE_ID> --region us-east-1
   sudo -u openclaw bash
   export HOME=/home/openclaw
   openclaw pairing approve telegram <CODE>

1. Bot is live! ✅

Teardown

# Using saved output:
./scripts/teardown.sh --from-output ./deploy-output.json --env-dir /path/to/workspace --yes

# Or by name (discovers via tags):
./scripts/teardown.sh --name starfish --region us-east-1 --env-dir /path/to/workspace --yes

Model Support

--model flag

Pass any model string — it goes directly into openclaw.json as model.primary:

# Default (MiniMax M2.1 on Bedrock — no API key needed, uses IAM role)
./scripts/deploy_minimal.sh --name starfish --region us-east-1

# Gemini Flash (needs GEMINI_API_KEY in .env.starfish)
./scripts/deploy_minimal.sh --name starfish --region us-east-1 \
  --model google/gemini-2.0-flash

AWS Bedrock

Bedrock IAM permissions (bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream) are always added to the instance role — regardless of which model you choose. This means any deployed instance can use Bedrock models out of the box via IAM role credentials (no API key needed).

Known Bedrock model IDs:

Note: Bedrock models must be enabled in your AWS account via the Bedrock console before use.

Gemini

If GEMINI_API_KEY is present in .env.starfish, it's stored in SSM and written to auth-profiles.json. If absent, it's simply skipped — no error.

.env.starfish

TELEGRAM_BOT_TOKEN=...     # Required — from @BotFather
GEMINI_API_KEY=...         # Optional — from aistudio.google.com (needed for Gemini models)

Architecture (Minimal)

┌─────────────────────────────────────────────────────┐
│                      VPC (10.50.0.0/16)             │
│  ┌───────────────────────────────────────────────┐  │
│  │           Public Subnet (10.50.0.0/24)        │  │
│  │  ┌─────────────────────────────────────────┐  │  │
│  │  │      EC2 t4g.medium (ARM64, 4GB)        │  │  │
│  │  │  ┌───────────────────────────────────┐  │  │  │
│  │  │  │       OpenClaw Gateway             │  │  │  │
│  │  │  │  • Node.js 22.14.0                 │  │  │  │
│  │  │  │  • Any model (Bedrock/Gemini/etc)   │  │  │  │
│  │  │  │  • Telegram channel                │  │  │  │
│  │  │  │  • Encrypted EBS (gp3, 20GB)       │  │  │  │
│  │  │  └───────────────────────────────────┘  │  │  │
│  │  └─────────────────────────────────────────┘  │  │
│  └───────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────┘
         ↑                              ↓
    SSM (no SSH/inbound)      Outbound HTTPS only

Critical Lessons Learned (22 Issues)

These are baked into the deploy script. See references/TROUBLESHOOTING.md for full details.

Instance Sizing

• t4g.medium (4GB) required — t4g.small (2GB) OOMs during npm install + gateway startup
• ARM64 — better price/performance than x86

Node.js

• Node 22+ required — OpenClaw 2026.x requires Node ≥22.12.0
• Official tarball install — NodeSource setup_22.x unreliable on AL2023 ARM64
• git required — OpenClaw npm install has git-based dependencies

npm

• Use openclaw@latest — bare openclaw may resolve to placeholder package (0.0.1)

Gateway Startup

• Use openclaw gateway run --allow-unconfigured — NOT gateway start (which tries systemctl --user and fails)
• Config file must be openclaw.json — not config.yaml
• gateway.mode: "local" — required or you get "Missing config" error
• gateway.auth.mode: "token" — "none" is invalid

Telegram

• plugins.entries.telegram.enabled: true — must be explicit
• dmPolicy: "pairing" — not "allowlist" (blocks everyone without user list)
• streamMode: "partial" — some models don't support streaming tools, use "off" as fallback

Model

• Gemini 2.0 Flash — recommended (free tier: 15 RPM, 1M tokens/day, supports tools)
• Auth profiles required — create auth-profiles.json in agent dir
• Bedrock format — amazon-bedrock/MODEL_ID (not bedrock/)
• Bedrock models need console enablement — Anthropic requires use case form

Systemd Service

• Simplified service file — removed ProtectHome, ReadWritePaths=/tmp/openclaw, PrivateTmp due to namespace issues
• Use NODE_OPTIONS="--max-old-space-size=1024" — helps prevent OOM

Security

• No inbound ports — SSM Session Manager only
• Secrets fetched from SSM at runtime — startup script fetches secrets each time the service starts; config files are ephemeral (rewritten on each start, never stored in repo or static images)
• Encrypted EBS — enabled by default in deploy script
• IMDSv2 required — HttpTokens=required

File Layout

scripts/
  deploy_minimal.sh        # One-shot deploy (VPC + EC2 + OpenClaw)
  teardown.sh              # Clean teardown of all resources
  setup_deployer_role.sh   # Create IAM role/user with minimum permissions
  preflight.sh             # Pre-deploy validation checks
  smoke_test.sh            # Post-deploy health verification

references/
  TROUBLESHOOTING.md   # All 22 issues + solutions
  config-templates/    # Ready-to-use config files
    gemini-flash.json  # OpenClaw config for Gemini Flash
    auth-profiles-gemini.json  # Auth profile template
    openclaw.service.txt  # Systemd unit file template
    startup.sh         # Startup script template

Config Templates

OpenClaw Config (gemini-flash.json)

See references/config-templates/gemini-flash.json — includes all required fields.

Auth Profiles (auth-profiles-gemini.json)

Create at ~/.openclaw/agents/main/agent/auth-profiles.json

Systemd Service (openclaw.service)

Simplified for reliability — security hardening removed due to namespace issues.

Cost Breakdown (~$30/mo)

Troubleshooting

"No API key found for amazon-bedrock"

Cause: OpenClaw needs models.providers config in openclaw.json with "auth": "aws-sdk". An auth-profiles.json entry alone is NOT sufficient.

Fix: Add to openclaw.json on the instance:

sudo -u openclaw bash
cd /home/openclaw/.openclaw
jq '.models = {
  "providers": {"amazon-bedrock": {"baseUrl": "https://bedrock-runtime.us-east-1.amazonaws.com", "api": "bedrock-converse-stream", "auth": "aws-sdk", "models": [{"id": "minimax.minimax-m2.1", "name": "MiniMax M2.1", "input": ["text"], "contextWindow": 128000, "maxTokens": 4096}]}},
  "bedrockDiscovery": {"enabled": true, "region": "us-east-1"}
}' openclaw.json > /tmp/oc.json && mv /tmp/oc.json openclaw.json
chown openclaw:openclaw openclaw.json
systemctl restart openclaw

"API rate limit reached" (Gemini)

Fix: Switch to Bedrock (default in current version) or redeploy with --model amazon-bedrock/minimax.minimax-m2.1.

Bedrock model returns errors

Cause: Model must be enabled in AWS Console → Bedrock → Model access. MiniMax models are auto-authorized; Anthropic/Meta models require use-case approval.

Bot doesn't respond after deploy

Fix: Add TELEGRAM_USER_ID to .env.starfish for auto-pairing, or use --pair-user <id>. Manual: openclaw pairing approve telegram <CODE> via SSM.

Safety Rules

• Never print secrets in logs
• Never open SSH/inbound ports; use SSM Session Manager only
• Use least-privilege IAM policies
• All resources tagged with Project=<name> and DeployId=<unique-id> for deterministic cleanup
• Encrypted EBS volumes always