OpenClaw Admin

🛡️ EMERGENCY SAFETY PROTOCOL

• NO OVERWRITES: Never use cat, redirection (>), or raw file writes to openclaw.json. Use a validated JSON utility script that loads-modifies-saves.
• SCHEMA VALIDITY: Before adding new keys (e.g., mcp, tools), verify against official schema. Never guess keys.
• PRE-BOUNCE VERIFICATION:

1. Validate JSON syntax (python -m json.tool).
2. Run openclaw doctor.
3. If any errors, ABORT the bounce.
• BACKUP INTEGRITY: Maintain a rotating history of up to 3 validated backups (openclaw.json.bak.1 to .bak.3). After a verified successful bounce, prune any backups older than 24 hours and maintain the most recent 3 verified snapshots.
• SECRET HYGIENE: Never write credential-like patterns (PAT, TOKEN, KEY) into any workspace file. Before removing any existing configuration that contains keys/credentials, ensure they are secured in the environment (~/.openclaw/credentials/.env) first.
• EXPLICIT AUTHORIZATION: I must never make changes to the openclaw.json or system configuration without first proposing the change, explaining the "Why," and obtaining explicit user permission.
• EXTERNAL SKILL AUDIT: Never install a skill from ClawHub or any external repository without first inspecting the source code. Look specifically for:
• Exfiltration of data/tokens to unknown URLs.
• Malicious/damaging system commands (e.g., recursive deletion, system modifications).
• Hardcoded or suspicious credential usage.

Only proceed with installation once the source has been audited as benign.

Execution Workflow

1. Clarify the target state. Ask what should change and what must stay untouched.
2. Select runtime scope first. Default profile unless isolation is explicitly needed:
• openclaw --dev ... → isolated dev state under ~/.openclaw-dev, gateway port 19001.
• openclaw --profile ... → isolated state under ~/.openclaw-.
• openclaw --container ... → target a named container for execution.
3. Route to the right command family. Use references/command-map.md for quick routing.
4. Expand subcommands before risky operations. Run openclaw --help for starred families; confirm flags before executing.
5. Prefer machine-readable output for automation. Use --json where available; parse and verify.
6. Verify outcomes explicitly. Check with openclaw status, openclaw health, openclaw doctor, or command-specific follow-up.

Safety Rules

• Require explicit user confirmation before reset, uninstall, destructive --force flows, or operations that remove stored provider keys.
• Prefer non-destructive diagnostics first: status, health, doctor, logs, security audit.
• Keep profile scope consistent across a workflow — never mix --dev and default in the same sequence.
• For gateway issues, diagnose before restart unless restart is explicitly requested.
• Keep the OpenAI HTTP endpoint on loopback/tailnet only — never expose to the public internet.

Triage Sequence

For any "OpenClaw not working" incident:

1. openclaw status
2. openclaw health
3. openclaw doctor
4. openclaw security audit (if config or provider connection settings may be misconfigured)
5. openclaw backup verify (if data integrity is suspected)
6. Check openclaw gateway ..., openclaw channels ..., or openclaw nodes ... based on where failure appears.
7. Escalate to targeted commands in references/command-map.md.

Architecture Overview

OpenClaw runs as a single Gateway (WebSocket, default 127.0.0.1:18789) that:

• Owns all messaging surfaces (WhatsApp/Telegram/Discord/Slack/Signal/iMessage/etc.)
• Hosts one or many agents, each with isolated workspace + auth + session store
• Exposes a typed WS API for CLI/app/automation clients
• Optionally serves an OpenAI-compatible HTTP surface at the same port

Clients (CLI, macOS app, web UI) connect over WebSocket. Nodes (macOS/iOS/Android/headless) also connect with role: node. The canvas UI is served at /__openclaw__/canvas/.

Workspace and Bootstrap Files

Each agent has a workspace directory (~/.openclaw/workspace by default) containing markdown files the agent reads on boot:

• Missing files get a placeholder marker injected at session start; execution continues.
• Size limits: bootstrapMaxChars (default 12 000 chars per file), bootstrapTotalMaxChars (default 60 000 total).
• openclaw setup --workspace recreates any missing defaults without overwriting existing ones.
• Keep the workspace in a private git repo for backup and recovery. Never commit ~/.openclaw/ state dirs.
• If ~/openclaw/ (old path) exists alongside ~/.openclaw/workspace, keep only one active workspace to avoid auth/session drift.

Agent System

Single-Agent Mode (default)

Out of the box: one agent, agentId = main, sessions keyed agent:main:.

Multi-Agent Setup

Each agent is a fully isolated brain:

• Own workspace (agents.defaults.workspace or per-agent override)
• Own agentDir (~/.openclaw/agents//agent) — holds auth-profiles.json, model registry, per-agent config
• Own session store (~/.openclaw/agents//sessions)
• Never reuse agentDir across agents — causes auth/session collisions.

Creating a new agent:

openclaw agents add coding
openclaw agents add work --workspace ~/.openclaw/workspace-work --non-interactive

Identity setup:

openclaw agents set-identity --agent main --name "OpenClaw" --emoji "🦞" --avatar avatars/oc.png
openclaw agents set-identity --workspace ~/.openclaw/workspace --from-identity   # reads IDENTITY.md

Listing and verifying:

openclaw agents list
openclaw agents list --bindings
openclaw agents bindings --agent work

Routing Bindings

Bindings route inbound messages to the correct agent. Most-specific match wins (priority order):

1. peer (exact DM/group id)
2. parentPeer (thread inheritance)
3. guildId + roles (Discord role routing)
4. guildId (Discord server)
5. teamId (Slack workspace)
6. Explicit accountId for a channel
7. accountId: "*" (channel-wide fallback, all accounts)
8. Default agent (agents.list[].default, else first entry, else main)

Managing bindings:

openclaw agents bind --agent work --bind telegram:ops --bind discord:guild-a
openclaw agents unbind --agent work --bind telegram:ops
openclaw agents unbind --agent work --all

Omitting --agent targets the current default. A channel-only binding (no accountId) is auto-upgraded to account-scoped when you later add an explicit accountId for the same channel+agent.

Config example (WhatsApp DM split):

{
  "agents": {
    "list": [
      { "id": "home", "default": true, "workspace": "~/.openclaw/workspace-home" },
      { "id": "work", "workspace": "~/.openclaw/workspace-work" }
    ]
  },
  "bindings": [
    { "agentId": "home", "match": { "channel": "whatsapp", "accountId": "personal" } },
    { "agentId": "work", "match": { "channel": "whatsapp", "accountId": "biz" } }
  ]
}

See references/multi-agent-recipes.md for full Discord, Telegram, and WhatsApp examples.

Sub-Agents

Sub-agents are background agent runs spawned by the main agent to handle tasks in parallel. They run in their own isolated session (agent::subagent:), receive only the instructions given to them (not full conversation history), and announce results back to the requester channel when done.

Key characteristics:

• Depth limit: currently flat (sub-agents cannot spawn their own sub-agents; this restriction is expected to be lifted in a future release)
• Restricted tool policies relative to the parent
• Auto-announce on completion (direct delivery first, falls back to queue routing, then exponential backoff)
• Tracked as background tasks; inspect/control via slash commands

Slash command control:

/subagents          # list all background sub-agent runs
/subagents spawn <agentId> <task> [--model <m>] [--thinking <level>]
/focus <target>     # bind current thread to a sub-agent session
/unfocus            # detach thread binding
/session idle       # inspect/update inactivity auto-unfocus
/session max-age    # control hard session age cap

When to use sub-agents vs sessions_send:

• Sub-agent (sessions_spawn): need the result now, in this conversation; parallel fan-out.
• sessions_send: fire-and-forget delegation; the other agent works independently and responds later.

Allowlist config (to let sub-agents target named agents):

{
  "agents": {
    "list": [
      {
        "id": "main",
        "subagents": {
          "allowAgents": ["coding", "research"]
        }
      }
    ]
  }
}

Running an Agent Turn

# Run via gateway (default)
openclaw agent --to +15555550123 --message "Status update" --deliver
openclaw agent --agent ops --message "Summarize logs" --thinking medium
openclaw agent --session-id 1234 --message "Summarize inbox" --json

# Run embedded (local, no gateway)
openclaw agent --agent ops --message "Run locally" --local

# Deliver to a different channel/account
openclaw agent --agent ops --message "Generate report" \
  --deliver --reply-channel slack --reply-to "#reports"

Key flags: --thinking , --verbose , --timeout .

Gateway mode falls back to embedded when the gateway request fails; --local forces embedded up front.

OpenAI-Compatible HTTP API

The Gateway can serve an OpenAI-compatible HTTP surface at the same port as WebSocket. Disabled by default.

Enable in ~/.openclaw/openclaw.json:

{
  "gateway": {
    "http": {
      "endpoints": {
        "chatCompletions": { "enabled": true }
      }
    }
  }
}

Endpoints (when enabled):

• POST /v1/chat/completions
• GET /v1/models / GET /v1/models/{id}
• POST /v1/embeddings
• POST /v1/responses

Model field = agent target:

• "openclaw" or "openclaw/default" → configured default agent
• "openclaw/" → specific agent (e.g. "openclaw/research")
• Legacy aliases: "openclaw:", "agent:"

Optional request headers:

• x-openclaw-model: — override backend model for the agent
• x-openclaw-agent-id: — compatibility agent override
• x-openclaw-session-key: — control session routing
• x-openclaw-message-channel: — set synthetic ingress channel context

Session behavior: stateless per request by default. If the request includes an OpenAI user string, a stable session key is derived from it so repeated calls share the same agent session.

Note: OpenClaw's gateway token is what any connected client uses to authenticate. Keep this endpoint on loopback or a private network (tailnet/VPN) — don't expose it directly to the public internet.

Examples:

# Non-streaming
curl -sS http://127.0.0.1:18789/v1/chat/completions \
  -H 'Authorization: Bearer YOUR_TOKEN' \
  -H 'Content-Type: application/json' \
  -d '{"model":"openclaw/default","messages":[{"role":"user","content":"hi"}]}'

# Streaming with backend model override
curl -N http://127.0.0.1:18789/v1/chat/completions \
  -H 'Authorization: Bearer YOUR_TOKEN' \
  -H 'Content-Type: application/json' \
  -H 'x-openclaw-model: openai/gpt-4o' \
  -d '{"model":"openclaw/research","stream":true,"messages":[{"role":"user","content":"hi"}]}'

# List agent targets
curl -sS http://127.0.0.1:18789/v1/models \
  -H 'Authorization: Bearer YOUR_TOKEN'

# Embeddings
curl -sS http://127.0.0.1:18789/v1/embeddings \
  -H 'Authorization: Bearer YOUR_TOKEN' \
  -H 'Content-Type: application/json' \
  -H 'x-openclaw-model: openai/text-embedding-3-small' \
  -d '{"model":"openclaw/default","input":["alpha","beta"]}'

Open WebUI quick setup:

• Base URL: http://127.0.0.1:18789/v1 (Docker on macOS: http://host.docker.internal:18789/v1)
• API key: your gateway bearer token
• Model: openclaw/default

Storing provider keys outside the config file (SecretRefs)

OpenClaw supports referencing provider keys via environment variables, files, or exec commands instead of writing plaintext values into openclaw.json. This is an OpenClaw configuration feature — the skill just explains how to set it up.

Set a reference via CLI:

openclaw config set providers.anthropic.key \
  --ref-provider env --ref-source ANTHROPIC_API_KEY

openclaw secrets configure       # interactive setup wizard
openclaw secrets reload          # reload without gateway restart
openclaw secrets audit           # check all references resolve correctly
openclaw secrets audit --check   # non-zero exit on failures (CI use)

Use ref mode during non-interactive onboard:

openclaw onboard --non-interactive \
  --auth-choice anthropic-api-key \
  --secret-input-mode ref \
  --anthropic-api-key secretref-env:ANTHROPIC_API_KEY

Setup and Onboarding

Interactive (recommended for first-time):

openclaw onboard

Non-interactive (VPS/headless):

openclaw onboard \
  --non-interactive \
  --auth-choice anthropic-api-key \
  --anthropic-api-key <key> \
  --gateway-bind loopback \
  --install-daemon \
  --daemon-runtime node \
  --node-manager pnpm \
  --skip-channels \
  --skip-search

Key --auth-choice values: anthropic-api-key, openai-api-key, openrouter-api-key, gemini-api-key, github-copilot, chutes, deepseek-api-key, custom-api-key, skip.

Key gateway bind modes: loopback (default), lan, tailnet, auto.

For Tailscale: --tailscale .

For remote gateway: --mode remote --remote-url --remote-token .

Models and Providers

Read references/models-and-providers.md for the full reference. It covers: model selection order, primary/fallback config, per-agent overrides, the model allowlist, custom providers, local models (Ollama/vLLM/LM Studio), multi-key rotation, how failover and cooldowns work, and sub-agent model config.

Quick orientation:

Model refs use provider/model format: anthropic/claude-opus-4-6, openai/gpt-5.4, ollama/llama3.3.

# Essential commands
openclaw models list                            # see configured models
openclaw models status                          # primary, fallbacks, auth overview
openclaw models set anthropic/claude-opus-4-6  # set primary model
openclaw models set-image openai/gpt-image-1   # set image model
openclaw models fallbacks add openrouter/auto   # add a fallback
openclaw models aliases add Opus anthropic/claude-opus-4-6
openclaw models auth login                      # add/re-auth a provider
openclaw models auth login-github-copilot       # GitHub Copilot device login
openclaw models scan                            # find free models on OpenRouter

Config: primary + fallback chain:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-opus-4-6",
        "fallbacks": [
          "openrouter/anthropic/claude-opus-4-6",
          "openai/gpt-5.4"
        ]
      }
    }
  }
}

Config: add a custom/third-party provider (any OpenAI-compatible API):

{
  "models": {
    "mode": "merge",
    "providers": {
      "my-provider": {
        "baseUrl": "https://api.example.com/v1",
        "apiKey": "${MY_PROVIDER_KEY}",
        "api": "openai-completions",
        "models": [
          { "id": "my-model", "name": "My Model" }
        ]
      }
    }
  }
}

Use api: "anthropic-messages" for Anthropic-compatible endpoints. After adding a provider: openclaw config validate, openclaw models list --provider my-provider, openclaw gateway restart.

Per-agent model (each agent can use a different model):

{
  "agents": {
    "list": [
      { "id": "main", "model": "anthropic/claude-opus-4-6" },
      { "id": "fast", "model": "anthropic/claude-haiku-4-5" }
    ]
  }
}

For built-in providers (Anthropic, OpenAI, Google, OpenRouter, GitHub Copilot, Mistral, xAI, DeepSeek, Ollama, and 20+ more) and their --auth-choice values, see references/models-and-providers.md.

Inference CLI (infer / capability)

Direct inference without a full agent run:

openclaw infer list                                    # list capabilities
openclaw infer model run --model anthropic/claude-sonnet-4-6 --prompt "hi"
openclaw infer image generate --prompt "a lobster in space"
openclaw infer image describe path/to/image.jpg
openclaw infer audio transcribe recording.mp3
openclaw infer tts convert --text "Hello" --voice nova
openclaw infer web search "latest OpenClaw release"
openclaw infer web fetch https://example.com
openclaw infer embedding create --input "embed this text"

Channels

openclaw channels list
openclaw channels status --probe            # live per-account probe
openclaw channels add --channel telegram --account alerts --token $TOKEN
openclaw channels add --channel discord --account work --token $DISCORD_TOKEN
openclaw channels remove --channel discord --account work --delete
openclaw channels login --channel whatsapp --account personal --verbose
openclaw channels logout --channel whatsapp --account personal
openclaw channels capabilities --channel telegram --account alerts
openclaw channels logs --channel whatsapp --lines 200

Supported channels: whatsapp, telegram, discord, googlechat, slack, signal, imessage, msteams, mattermost (plugin), and many more via plugins.

DM security policies: dmPolicy: "pairing" (unknown senders get a code), dmPolicy: "allowlist" (explicit list only), dmPolicy: "open".

Resources

Read the appropriate reference file before answering deep questions in these areas:

Live docs: openclaw docs [query] or https://docs.openclaw.ai