agent-spawner（特工产卵者）

Agent Spawner

Deploy a new OpenClaw agent conversationally. Official install, carry over config from the current agent. User never edits a file.

1. Read Current Config (silent)

cat ~/.openclaw/openclaw.json
cat ~/.openclaw/.env 2>/dev/null
env | grep -iE 'API_KEY|TOKEN'
ls ~/.openclaw/extensions/
ls <workspace>/skills/

Identify:

• Provider: check auth.profiles in config — could be Anthropic, OpenAI, Gemini, custom, etc.
• API key: from env var or config (e.g. ANTHROPIC_API_KEY, GEMINI_API_KEY, OPENAI_API_KEY)
• Model: from agents.defaults.model
• Tool keys: anything in tools.* (search APIs, etc.)
• Plugins: plugins.installs — names and npm specs
• Skills: run openclaw skills list to see what's bundled vs workspace-only. Only carry over non-bundled skills.

2. Ask

1. "Where should I deploy it?" — Docker (local or remote SSH) or bare metal?
2. "Name?" — for container. Generate one if they don't care.
3. "Anything special?" — purpose, constraints. Optional.

Don't ask about keys, plugins, skills, ports, or config. Carry everything over, use defaults.

3. Confirm Plan

After gathering answers, present the full plan before doing anything. Show everything in one summary:

Here's the plan:

📦 Deploy: Docker on <target>
📛 Name: <agent-name>
🌐 Port: <port>

Carrying over from current agent:
  ✅ Provider: Anthropic (API key)
  ✅ Model: anthropic/claude-sonnet-4-20250514
  ✅ Brave Search API key
  ✅ Plugins: openclaw-agent-reach
  ✅ Skills: agent-spawner, weather
  ✅ Heartbeat: 30m

The new agent will bootstrap its own identity on first message.

Good to go?

Only list items that actually exist. Wait for explicit confirmation before proceeding. If the user wants changes, adjust and re-confirm.

4. Deploy

Docker

git clone https://github.com/openclaw/openclaw.git <agent-name>
cd <agent-name>

Set env and run non-interactive onboard. Match the provider detected in step 1:

export OPENCLAW_IMAGE=alpine/openclaw:latest
export OPENCLAW_CONFIG_DIR=~/.openclaw-<agent-name>
export OPENCLAW_WORKSPACE_DIR=~/.openclaw-<agent-name>/workspace
export OPENCLAW_GATEWAY_PORT=<unused port, default 18789>
export OPENCLAW_GATEWAY_BIND=lan

mkdir -p $OPENCLAW_CONFIG_DIR/workspace

Onboard flags vary by provider. Use the matching --auth-choice and key flag:

docker compose run --rm openclaw-cli onboard --non-interactive --accept-risk \
  --mode local \
  --auth-choice <detected> \
  --<provider>-api-key "$API_KEY" \
  --gateway-port 18789 \
  --gateway-bind lan \
  --skip-skills

docker compose up -d openclaw-gateway

Official compose uses bind mounts — host user owns files, no permission issues.

Onboard error about gateway connection is expected (not running yet). Config is written.

Bare metal

curl -fsSL https://openclaw.ai/install.sh | bash -s -- --no-onboard

openclaw onboard --non-interactive --accept-risk \
  --mode local \
  --auth-choice <detected> \
  --<provider>-api-key "$API_KEY" \
  --gateway-port 18789 \
  --gateway-bind lan \
  --install-daemon \
  --daemon-runtime node \
  --skip-skills

5. Patch Running Agent

CLI alias:

• Docker: OC="docker compose exec openclaw-gateway node /app/openclaw.mjs"
• Bare metal: OC="openclaw"

Config (only patch what the current agent actually has):

$OC config set agents.defaults.model "<model>"
$OC config set agents.defaults.heartbeat.every "30m"
# Tool keys — only if they exist in current config
$OC config set tools.web.search.apiKey "<key>"

Plugins (from plugins.installs in current config):

$OC plugins install <npm-spec>
# Repeat for each plugin

Skills (copy workspace skills):

# Docker
docker cp <source-workspace>/skills/ <container>:/home/node/.openclaw/workspace/skills/
# Bare metal
cp -r <source-workspace>/skills/ ~/.openclaw/workspace/skills/

Restart:

docker compose restart openclaw-gateway  # Docker
openclaw gateway restart                 # bare metal

6. Hand Off

Read the gateway token:

grep -A1 '"token"' $OPENCLAW_CONFIG_DIR/openclaw.json

Tell the user:

• URL: http://<host>:<port>/
• Token: (from config — onboard auto-generates one)
• "Say hello — it'll bootstrap itself."

Notes

• openclaw not in PATH inside Docker. Use node /app/openclaw.mjs.
• --accept-risk required for non-interactive onboard.
• alpine/openclaw:latest — pre-built official image.
• Don't use named Docker volumes — root ownership issues. Official compose uses bind mounts.
• Multiple agents on same host: use different OPENCLAW_CONFIG_DIR and OPENCLAW_GATEWAY_PORT.
• Plugins and skills persist in ~/.openclaw/ volume (extensions/ and workspace/skills/).
• SSH keys, git config, apt packages are ephemeral — not in the volume, by design.