Half the OpenClaw cheat sheets online list commands the CLI doesn’t have. Type openclaw tts, openclaw image or openclaw start on the current release and you get OpenClaw does not know the command, because those features live under openclaw infer and openclaw gateway now. This OpenClaw cheat sheet sticks to the commands that actually answer.
You get the OpenClaw commands for install, the Gateway service, config, models, agents, channels, cron and backups, plus the errors you’ll hit. Each row was rerun on Ubuntu 24.04 with OpenClaw 2026.11.3 in November 2026, or checked against that build’s own --help output, and the few outputs kept from the first 2026.11.2 run say so.
OpenClaw cheat sheet (quick reference)
The tables are grouped by job, so scan for the task rather than the command name. OpenClaw ships more than 70 top-level commands and most of them take subcommands, which is why a complete dump is useless at the terminal. These are the ones you will reach for every week, and the full list is one openclaw --help away. Anything marked with a placeholder like <id> comes from the output of a list command in the same table.
Install and update
| Command | What it does |
|---|---|
curl -fsSL https://openclaw.ai/install.sh | bash -s -- --no-onboard | Install Node 24 and OpenClaw, skip the setup wizard |
openclaw --version | Print the version and build hash |
openclaw update status | Install kind, channel and the latest npm release |
openclaw update --dry-run | Preview an update without changing anything |
openclaw update --dry-run --channel beta | Preview a move to the beta channel |
npm install -g openclaw@latest --allow-scripts=openclaw | Upgrade with npm 11.16+, then run openclaw gateway restart |
openclaw completion --shell zsh | Print a completion script (zsh, bash, fish, powershell) |
Onboard and set up
| Command | What it does |
|---|---|
openclaw onboard | Guided setup for auth, model, Gateway and channels |
openclaw onboard --install-daemon | Same, and install the Gateway as a systemd user service |
openclaw onboard --non-interactive --accept-risk --auth-choice ollama --install-daemon | Scripted setup against a local Ollama |
openclaw setup --baseline | Create the baseline config, workspace and session folders |
openclaw configure | Interactive edit of credentials, channels and defaults |
openclaw dashboard --no-open | Print the Control UI URL instead of opening a browser |
Gateway service
| Command | What it does |
|---|---|
openclaw gateway status | Service state, port, bind address and versions |
openclaw gateway start | Start the installed service (systemd, launchd, schtasks) |
openclaw gateway stop --force | Stop the service (scripts need –force) |
openclaw gateway restart | Restart the service |
openclaw gateway run --force | Run in the foreground, killing anything on the port |
openclaw --dev gateway | Isolated dev Gateway on port 19001 |
openclaw gateway usage-cost --days 7 | Token cost summary from session logs |
openclaw health | Event loop, agents and session store health |
openclaw logs --follow | Tail the Gateway log |
openclaw logs --limit 15 --plain | Last 15 lines, no colour |
Status and diagnostics
| Command | What it does |
|---|---|
openclaw status | One-screen overview of Gateway, agents and sessions |
openclaw status --all | Full read-only diagnosis you can paste into an issue |
openclaw status --deep | Also probe channel credentials |
openclaw doctor --fix | Apply recommended repairs |
openclaw doctor --lint | Read-only checks as JSON findings |
openclaw security audit | Flag risky config (small models, exposure, proxies) |
openclaw secrets audit --check | Exit non-zero if plaintext secrets remain |
Config
| Command | What it does |
|---|---|
openclaw config file | Print the active config path |
openclaw config get agents.defaults.model.primary | Read one value by dot path |
openclaw config set agents.defaults.thinkingDefault low | Write one value |
openclaw config unset agents.defaults.thinkingDefault | Remove a value |
openclaw config validate | Check the file against the schema |
openclaw config patch --file ./p.json5 --dry-run | Preview a multi-key merge |
openclaw config schema | Print the JSON schema for openclaw.json |
openclaw --profile work config file | Use an isolated profile under ~/.openclaw-work |
Models and aliases
| Command | What it does |
|---|---|
openclaw models list | Configured models with context size and auth |
openclaw models list --all | Full model catalog |
openclaw models status | Default, fallbacks, aliases and auth health |
openclaw models set <model-or-alias> | Change the default model |
openclaw models aliases add fast ollama/qwen3:1.7b | Create an alias |
openclaw models aliases list | List aliases |
openclaw models aliases remove fast | Delete an alias |
openclaw models fallbacks add <model> | Add a fallback model |
Agents, sessions and chat
| Command | What it does |
|---|---|
openclaw agent --agent main -m "..." | Run one agent turn through the Gateway |
openclaw agent --local --agent main -m "..." | Embedded run (Gateway must be stopped) |
openclaw agents add research --non-interactive --workspace ~/research-ws | Add an isolated agent |
openclaw agents list | Agents with workspace, model and routing |
openclaw agents delete research --force | Remove an agent without a prompt |
openclaw sessions list | Stored sessions with model and token use |
openclaw sessions tail --follow | Stream a running session’s events |
openclaw chat | Local terminal UI (alias for tui –local) |
openclaw tui | Terminal UI connected to the Gateway |
openclaw resume | Reopen a recent Gateway session in the TUI |
Channels and messages
| Command | What it does |
|---|---|
openclaw channels list --all | Every channel plugin and its install state |
openclaw channels add --channel telegram | Add a channel account with guided prompts |
openclaw channels status --probe | Check channel credentials live |
openclaw channels remove --channel telegram --delete | Drop a channel’s config, no prompt |
openclaw message send --channel telegram --target @mychat -m "Hi" | Send a message through a channel |
openclaw pairing list telegram | Pending DM pairing requests on a channel |
openclaw pairing approve telegram <code> | Approve a pairing code |
Inference, images and speech
| Command | What it does |
|---|---|
openclaw infer list | Every inference capability and its transport |
openclaw infer model run --prompt "..." | One-shot model call, no agent tools |
openclaw infer image providers | Image providers and their models |
openclaw infer image generate --prompt "..." --output a.png | Generate an image (needs an image model configured) |
openclaw infer tts providers | Speech providers |
openclaw infer tts convert --text "..." --output a.mp3 | Text to speech (needs a TTS provider) |
openclaw infer web search --query "..." | Web search (needs a search provider) |
Cron automations
| Command | What it does |
|---|---|
openclaw cron list | Jobs with schedule, next run and delivery |
openclaw cron add --name disk-check --every 1h --command 'df -h /' | Run a shell command hourly |
openclaw cron add --name brief --cron "0 7 * * *" --message "..." | Daily agent prompt at 07:00 |
openclaw cron run <id> --wait | Run a job now and wait for it |
openclaw cron rm <id> | Delete a job |
Plugins, skills and MCP
| Command | What it does |
|---|---|
openclaw plugins list | Stock and installed plugins, enabled or not |
openclaw plugins search <query> | Search the plugin registry |
openclaw plugins install <spec> | Install from npm, a path or an archive |
openclaw plugins enable <id> | Enable a stock plugin |
openclaw skills list --eligible | Skills ready to use |
openclaw skills check --agent main | Missing requirements per skill |
openclaw mcp list | OpenClaw-managed MCP servers |
openclaw mcp set <name> '<json>' | Add an MCP server (stop the Gateway first) |
Backup, reset and uninstall
| Command | What it does |
|---|---|
openclaw backup create --verify --output ~/backups | Archive ~/.openclaw into an existing directory |
openclaw backup verify <archive> | Check an existing archive |
openclaw backup restore <archive> --target ./restored | Restore into an empty dir, then activate it |
openclaw reset --scope config --dry-run --non-interactive --yes | Preview a config reset |
openclaw uninstall --all --dry-run --non-interactive --yes | Preview a full uninstall |
Two global flags change where every command above reads and writes. --profile <name> isolates state under ~/.openclaw-<name>, and --dev uses ~/.openclaw-dev with the Gateway on port 19001, so a test setup never touches your real agent.
Install OpenClaw and run onboarding
OpenClaw needs Node 24.16 or newer on the 24 line, or Node 26.1 and up, and the docs strongly recommend 24.22.1+ or 26.12.1+ for security fixes. The official installer handles that for you on Linux: it adds build tools, pulls Node 24 from NodeSource when Node is missing, and installs the npm package. Pass --no-onboard so it doesn’t drop you into the wizard before you’ve picked a model provider:
curl -fsSL https://openclaw.ai/install.sh | bash -s -- --no-onboard
On a fresh Ubuntu 24.04 VM the interesting part is the npm prefix. A non-root user can’t write to /usr, so the installer switches npm to a per-user prefix and warns you off sudo npm for future updates:
✓ Node.js v24.22.3 installed
· Active Node.js: v24.22.3 (/usr/bin/node)
· Active npm: 11.19.1 (/usr/bin/npm)
[2/3] Installing OpenClaw
✓ Git already installed
! npm global prefix is not writable: /usr
! The installer will switch npm's user prefix to /home/joshua/.npm-global; npm normally writes that setting to ~/.npmrc.
· Configuring npm for user-local installs
! Avoid sudo npm i -g for future OpenClaw updates; use npm i -g openclaw@latest so npm keeps using this user prefix instead of a different global prefix.
✓ npm configured for user installs
· Installing OpenClaw v2026.11.3
· Installing OpenClaw package
✓ OpenClaw npm package installed
· Published openclaw bin link at /home/joshua/.npm-global/bin/openclaw
✓ OpenClaw installed
That prefix puts the binary in ~/.npm-global/bin, which a new login shell picks up. If openclaw is not found in the session you installed from, open a new shell before blaming the install. The warning about sudo npm matters later too, because a root-owned global install next to the user one leaves you with two copies and openclaw update refreshing the wrong one. Confirm the build.
openclaw --version
The version string carries the git hash, which is what bug reports ask for.
OpenClaw 2026.11.3 (847611d)
Onboarding is where most first runs go sideways. The lab used a local Ollama container serving qwen3:1.7b so no API key was involved, and the first scripted attempt skipped --install-daemon. Non-interactive onboarding then waits for a Gateway that nothing started, and fails with exit code 1:
openclaw onboard --non-interactive --accept-risk --auth-choice ollama --flow quickstart --skip-channels --skip-skills --skip-search --skip-ui
Config and workspace get written, but the health check at the end cannot reach port 18789. The fix line at the bottom names all three ways out. This output is from the first run on 2026.11.2, on a VM where no Gateway was running yet.
Default Ollama model: qwen3:1.7b
Workspace OK: ~/.openclaw/workspace
Sessions OK: ~/.openclaw/agents/main/sessions
Updated config: ~/.openclaw/openclaw.json
Backup: ~/.openclaw/openclaw.json.bak
Gateway did not become reachable at ws://127.0.0.1:18789.
Classification: not-listening
Last check: connect ECONNREFUSED 127.0.0.1:18789
Non-interactive local setup only waits for an already-running gateway unless you pass `--install-daemon` to `openclaw onboard`.
Fix: start `openclaw gateway run`, re-run `openclaw onboard --install-daemon`, or use `openclaw onboard --skip-health`.
Add --install-daemon and onboarding writes a systemd user unit, turns on lingering so the Gateway survives logout, and starts it. That’s the version to put in a provisioning script:
openclaw onboard --non-interactive --accept-risk --auth-choice ollama --install-daemon --skip-channels --skip-skills --skip-search --skip-ui
This time it exits 0 and ends with a health report. The event loop line reads ok, and the Brave tip at the bottom only matters if you plan to give the agent web search:
Default Ollama model: qwen3:1.7b
[state/agent-db] agent database clean-close receipt: written
Workspace OK: ~/.openclaw/workspace
Sessions OK: ~/.openclaw/agents/main/sessions
Updated config: ~/.openclaw/openclaw.json
Backup: ~/.openclaw/openclaw.json.bak
Installed systemd service: /home/joshua/.config/systemd/user/openclaw-gateway.service
Enabled systemd lingering for joshua.
Gateway event loop: ok max=2ms p99=0ms util=0.008 cpu=1.408
Gateway check duration: 126ms
Agents: main (default)
Heartbeat interval: 30m (main)
Session store (main): /home/joshua/.openclaw/agents/main/agent/openclaw-agent.sqlite (0 entries)
Tip: run `openclaw configure --section web` to store your Brave API key for web_search. Docs: https://docs.openclaw.ai/tools/web
Machine size matters less than it did. On 2026.11.2 a 4 vCPU VM printed Gateway did not become reachable with Last check: timeout even though the service was running, and a restart took 41.6 seconds to log gateway ready. The 2026.11.3 rerun above was on a 4 vCPU VM and passed first time, though a second onboard on the same VM still logged a brief degraded event loop. That fits the release’s note about cutting Gateway startup overhead. If an older build or a loaded box still gives you that timeout, wait a minute and run openclaw gateway status before re-running onboarding. The service is a user unit, so manage it with systemctl --user, never with sudo.
Manage the OpenClaw Gateway service
Everything except embedded runs talks to the Gateway, a WebSocket server on 127.0.0.1:18789. When a command hangs or times out, gateway status is the first thing to run because it shows the unit file, the bind address and whether the CLI and the running Gateway are the same version:
openclaw gateway status
Read three lines in the output. Connectivity check: ok means the CLI reached the socket, bind=loopback means nothing outside the box can, and the two version lines should match. A mismatch after an update means the old process is still running, so restart it before you trust any behaviour you see.
Service: systemd user (enabled)
File logs: /tmp/openclaw/openclaw-2026-11-14.log
Service file: ~/.config/systemd/user/openclaw-gateway.service
Service env: OPENCLAW_GATEWAY_PORT=18789
Config (cli): ~/.openclaw/openclaw.json
Config (service): ~/.openclaw/openclaw.json
Gateway: bind=loopback (127.0.0.1), port=18789 (service args)
Dashboard: http://127.0.0.1:18789/
Check note: Loopback-only gateway; only local clients can connect.
CLI version: 2026.11.3 (~/.npm-global/bin/openclaw)
Gateway version: 2026.11.3
Runtime: running (pid 7583, state active, sub running, last exit 0, reason 0)
Connectivity check: ok
Capability: connected-no-operator-scope
Listening: 127.0.0.1:18789
Stop, start and restart are subcommands of gateway, not top-level commands. They drive whichever service manager onboarding installed (systemd here, launchd on macOS, scheduled tasks on Windows):
openclaw gateway restart
The restart returns as soon as systemd accepts it, not when the Gateway is ready.
Restarted systemd service: openclaw-gateway.service
Stopping is guarded, restarting isn’t. Its help describes --force as allowing a stop from a non-interactive shell, and a plain openclaw gateway stop run from the lab script exited 1 with This stops the operator's running gateway service ... re-run with --force if you really mean it. The CLI assumes you’d rather test against a --dev Gateway than take down the one your channels depend on. With --force it stops the unit, and gateway status then reads Connectivity check: failed and Runtime: stopped, which is the window to run commands that need the state offline:

After a start or restart, check that the Gateway answers before you send it work. Commands that land in the warm-up window fail to connect, and a script that does not wait will report a broken install that is really just a slow one.
openclaw health
A healthy Gateway reports a quiet event loop and a session store path.
Gateway event loop: ok max=10ms p99=0ms util=0.002 cpu=0.002
Gateway check duration: 198ms
Agents: main (default)
Heartbeat interval: 30m (main)
Session store (main): /home/joshua/.openclaw/agents/main/agent/openclaw-agent.sqlite (0 entries)
For a wider view, openclaw status prints one table covering the Gateway, service, agents, telemetry, memory and backups. It also tells you the heartbeat has nowhere to deliver until a channel exists, which is expected on a fresh install with no Telegram or Slack account added.

Two rows in that table are worth reading every time you run it. Gateway shows reachability and the auth mode, so a reachable 2.99s on a local socket tells you the box is under load. Update shows whether you installed from npm or a git checkout, and that decides how openclaw update behaves, because a git install rebases and rebuilds while an npm install just swaps the package.
Read and change OpenClaw config from the CLI
All settings live in one JSON file, ~/.openclaw/openclaw.json by default. You don’t need an editor for routine changes: config get, set and unset work on dot paths and validate before writing. Find the file first, since profiles and environment variables move it:
openclaw config file
On the lab box it is the default path.
/home/joshua/.openclaw/openclaw.json
The capture below is a full round trip on the default thinking level. Watch the message after set. This key applies live, so there is no Gateway restart to schedule, and validate at the end proves the file still matches the schema after both writes.

Not every key hot-reloads, so read that last line on each set rather than assuming. The path printed by config file also changes with the flags and variables below, which is how you keep a scratch setup away from your real agent:
openclaw --profile work config fileprinted/home/joshua/.openclaw-work/openclaw.json.OPENCLAW_CONFIG_PATH=/tmp/alt.json openclaw config fileprinted/tmp/alt.json.openclaw --dev config fileprinted/home/joshua/.openclaw-dev/openclaw.json.
One thing other sheets get wrong here: there’s no config show or config edit. The subcommands are file, get, patch, schema, set, unset and validate, and openclaw config with no subcommand starts guided setup. For several keys at once, write them to a JSON5 file and use config patch --file ./p.json5 --dry-run before applying it for real.
Switch models and add aliases
Model IDs take the form provider/model, so the Ollama model in the lab is ollama/qwen3:1.7b. There are no built-in short names to memorize: models aliases list on a fresh install returns Aliases (0). Every alias is one you create, which is why copying alias names from another sheet fails.
openclaw models list
The list shows input types, the context window in use against the maximum, and whether auth resolved. The 33k/41k figure matters later, because every agent turn has to fit its system prompt inside that window.
Model Input Ctx Local Auth Tags
ollama/qwen3:1.7b text 33k/41k yes yes default
The alias workflow is four steps, and the capture runs all of them in order.
- Check what exists with
openclaw models aliases list. - Create the alias with
openclaw models aliases add fast ollama/qwen3:1.7b. - Make it the default with
openclaw models set fast, which stores the full ID. - Remove it with
openclaw models aliases remove fastwhen you are done.
Notice that models set fast printed Default model: ollama/qwen3:1.7b, not the alias. The config keeps the resolved ID, so deleting the alias afterwards does not break the default.

Each of those writes backs up the previous file to openclaw.json.bak first, so a bad change is one copy away from undone. openclaw models status then shows the default, fallbacks, image model, aliases and where each provider’s credentials come from, which is the quickest way to see why an agent picked the model it did.
Run agent turns and one-shot inference
openclaw agent sends one message to an agent through the Gateway, with its tools, memory and workspace. openclaw infer model run is the bare model call with none of that, so it’s the right test when you only want to know whether the provider answers.
time openclaw infer model run --prompt 'Say hello in five words.' --thinking off
It ran locally against Ollama with no GPU. The output below is from the first run of this sheet on a 16 vCPU VM, where it answered in under 9 seconds; the 2026.11.3 rerun on 4 vCPUs printed the same lines in 17.3 seconds.
model.run via local
provider: ollama
model: qwen3:1.7b
outputs: 1
Hello, world!
real 0m8.965s
user 0m4.985s
sys 0m0.548s
The same model behind a full agent turn is a different story. The agent sends its system prompt, tool schemas and workspace bootstrap files with every message, so even a one-line question costs thousands of tokens of prompt processing:
time openclaw agent --agent main --thinking off --timeout 1200 -m 'Reply with one short sentence: what is the capital of Kenya?'
On the 16 vCPU VM the answer was right, and it took 1 minute 43 seconds to arrive, about eleven times the bare model call.
The capital of Kenya is Nairobi.
real 1m43.175s
user 0m1.185s
sys 0m0.149s
openclaw sessions list explains the gap. That single question filled 13k of the model’s 33k context window, 39% before the conversation had started:

The 4 vCPU rerun on 2026.11.3 is where it breaks, and --timeout 1200 didn’t save it. After 10 minutes 14 seconds the provider’s own idle timer fired first and the turn exited 1:
The model did not produce a response before the model idle timeout. Please try again, or increase `models.providers.<id>.timeoutSeconds` for slow local or self-hosted providers. If `agents.defaults.timeoutSeconds` or a run-specific timeout is lower, raise that ceiling too; provider timeouts cannot extend a model attempt budget.
real 10m14.735s
So there are two clocks. --timeout caps the whole agent run, while models.providers.ollama.timeoutSeconds caps how long OpenClaw waits on the model, and the lower one wins. Raising the provider value to 1800 got the same question answered in 15 minutes 28 seconds. Set it back once you move to a faster model:
openclaw config set models.providers.ollama.timeoutSeconds 1800
openclaw config unset models.providers.ollama.timeoutSeconds
Fifteen minutes for one sentence is the real lesson. Judge OpenClaw’s speed on a hosted or GPU-backed model rather than a small local one on a few CPU cores, and use infer model run when you only want to prove the provider works.
Extra agents get their own workspace, auth and session store, which keeps a research agent’s memory out of your main one. Adding one non-interactively takes a name, a workspace and a model:
openclaw agents add research --non-interactive --workspace ~/research-ws --model ollama/qwen3:1.7b
The new agent gets its own directory under ~/.openclaw/agents, while its workspace sits wherever you pointed it.
Updated config: ~/.openclaw/openclaw.json
Backup: ~/.openclaw/openclaw.json.bak
Agent: research
Workspace: ~/research-ws
Agent dir: ~/.openclaw/agents/research/agent
Model: ollama/qwen3:1.7b
Once a second agent exists, commands that used to infer the agent start asking. openclaw skills check refused to guess in the lab until it got --agent main, so add that flag to any script that ran fine with one agent. The sessions commands here stop at reading that store, and if you split your time between the two CLIs, the OpenCode CLI cheat sheet shows how opencode forks, deletes and exports its own sessions.
Check which OpenClaw commands don’t exist
Older sheets and some hosting-provider guides were written against earlier builds, and a few commands they list now fail. The CLI moved media and speech under infer and service control under gateway, so the old top-level names fall through to the unknown-command handler. All four of these fail on OpenClaw 2026.11.3, with the same wording 2026.11.2 gave.

Every row in the table was typed into the same build, and the middle column is what the CLI said back. The right column is the command that does the job today.
| You type | 2026.11.3 says | Use instead |
|---|---|---|
openclaw start | does not know the command “start” | openclaw gateway start |
openclaw restart | does not know the command “restart” | openclaw gateway restart |
openclaw image | does not know the command “image” | openclaw infer image generate |
openclaw tts | does not know the command “tts” | openclaw infer tts convert |
openclaw channel list | Did you mean this? openclaw channels | openclaw channels list |
openclaw config show | config has no command “show” | openclaw config get <path> |
openclaw channels add --channel telegram --token <token> | Works: Added Telegram account “default” | Keep it, even though --token is missing from the Options list in --help |
openclaw chat | Works: alias for tui --local | Keep it |
The chat row is the one that confuses people. It does exist, but it opens the local terminal UI against the embedded runtime, not a Gateway session, so you won’t see the conversation your phone is having with the agent. Use openclaw tui when you want the same session your Telegram or Slack messages land in. When an unknown command comes back, the CLI’s own hint (Plugin command? openclaw plugins list) is worth following, since plugins can add top-level commands of their own.
Schedule OpenClaw automations with cron
openclaw cron (also reachable as openclaw automations) schedules work inside the Gateway, not in the system crontab. A job can send a prompt to an agent or run a shell command, on an interval, a cron expression or once with --at. Here’s an hourly disk check:
openclaw cron add --name disk-check --every 1h --command 'df -h /' --agent main
The job is created, but read the deliveryPreview at the bottom of the JSON it prints. With no channel configured, the result has nowhere to go. The JSON below is from the 2026.11.2 run on a box with no channels; on the 2026.11.3 rerun, where Telegram was already added, the same field said Delivering to Telegram requires target <chatId> instead, so a channel alone is not enough without a target.
"sessionTarget": "isolated",
"wakeMode": "now",
"payload": {
"kind": "command",
"argv": [
"sh",
"-lc",
"df -h /"
]
},
"delivery": {
"mode": "announce",
"channel": "last"
},
"deliveryPreview": {
"label": "announce -> last",
"detail": "last -> no route, will be blocked: Channel is required (no configured channels detected). Run openclaw channels add to configure one, or pass --channel <channel> after enabling a channel. Use openclaw channels list --all to see available channel ids. Set delivery.channel explicitly or use a main session with a previous channel. Configure a channel and delivery target, or use delivery:{mode:\"none\"} for no automatic delivery."
}
Nothing fails at creation time, so a job like this sits in cron list looking healthy and never reports. Either add a channel first or create the job with --no-deliver. The listing also shows two jobs you didn’t create: the 30-minute heartbeat for the main agent and a nightly memory job at 03:00 from the memory-core plugin. The Declaration column names the owner of each built-in job (heartbeat:main, memory-core:...) and shows a dash for the jobs you added, which makes your own work easy to pick out. The table is wider than 120 columns, so the capture is cut to fit.

Agent prompts take --message, not -m. openclaw agent accepts the short form, so it is easy to carry it over, but cron add rejected it in the lab with OpenClaw does not recognize option "-m" and created nothing. Remove a job with openclaw cron rm and the ID from the first column, or trigger one with cron run <id> --wait to test it before the schedule fires.
Audit security and back up OpenClaw state
An agent with shell, browser and web tools is root-adjacent on whatever box it runs on, so run the audit before you connect any chat channel. On the lab setup it flagged one critical finding that’s specific to running small local models:
openclaw security audit
The 1.7B model counts as small, and with web tools allowed it will act on whatever a fetched page tells it. That is the prompt-injection path the audit is worried about.

Before you connect a chat channel to a box like this, run the checks in order.
- Apply the fix the audit prints for small models: deny
group:webandbrowserfor that model and turn on the sandbox. - Run
openclaw secrets audit --check. In the lab it exited 1 withPLAINTEXT_FOUNDforgateway.auth.token. - Move that token into a SecretRef with
openclaw secrets configure, then rerun the audit until it exits 0.
The plaintext token is the one that bites people, because any agent tool that can read files can read openclaw.json too. openclaw doctor --lint flags the same path, so whichever check you run first, the fix is the same.
Backups are one command, and --verify reopens the archive after writing it. Create the target directory first: on a VM where ~/backups didn’t exist, the same command wrote the tarball as a single file named ~/backups and reported success, so the next backup verify ~/backups/*.tar.gz found nothing.
mkdir -p ~/backups
openclaw backup create --verify --output ~/backups
The state directory always goes in, and the ~/research-ws workspace of the agent added earlier got its own entry because it lives outside ~/.openclaw. Regenerable caches are skipped, and 2026.11.3 also skips volatile files such as the restart log, which it reports twice:
Backup skipped 1 volatile file (live sessions, cron logs, queues, managed runtime paths, sockets, pid/tmp).
Backup archive: /home/joshua/backups/2026-11-14T05-59-19.384+00-00-openclaw-backup.tar.gz
Included 2 paths:
- state: ~/.openclaw
- workspace: ~/research-ws
Skipped 6 paths:
- workspace: ~/.openclaw/workspace (covered by ~/.openclaw)
- agent: ~/.openclaw/agents/main/agent (covered by ~/.openclaw)
- agent: ~/.openclaw/agents/research/agent (covered by ~/.openclaw)
- plugin skills: ~/.openclaw/plugin-skills (regenerable)
- managed state: ~/.openclaw/tmp (regenerable)
- entry: /home/joshua/.openclaw/logs/gateway-restart.log (volatile)
Created /home/joshua/backups/2026-11-14T05-59-19.384+00-00-openclaw-backup.tar.gz
Skipped 1 volatile file (live sessions, cron logs, queues, managed runtime paths, sockets, pid/tmp).
Archive verification: passed
Restoring is a two-stage job. openclaw backup restore <archive> --target ./restored only unpacks into an empty directory; the output then tells you to stop the Gateway, move the restored tree into place or point OPENCLAW_STATE_DIR at it, and run openclaw doctor. Run openclaw backup verify on the archive first, which in the lab reported 102 entries scanned and the SQLite inventory verified.
A workspace you moved outside ~/.openclaw, like the ~/research-ws agent above, is a separate path to watch. Check the “Included” list after adding agents so a restore doesn’t come back missing one. For an always-on agent box, the Mac mini picks for 24/7 AI agents cover the external SSD and UPS that make those backups and the Gateway survive a power cut.
Fix common OpenClaw errors
These came up on the lab box, word for word. Most trace back to the Gateway owning state that a command wanted to touch directly.
Gateway did not become reachable at ws://127.0.0.1:18789
Non-interactive onboarding ran without --install-daemon, so no Gateway was started and the health check had nothing to talk to. Re-run it with that flag, start one in the foreground with openclaw gateway run, or skip the check with --skip-health if your Gateway runs on another host. If you did pass --install-daemon and still see it with Last check: timeout, the box is probably just slow, so give the service a minute and check openclaw gateway status before you change anything.
A Gateway is running for this state directory
openclaw agent --local refuses to run next to a live Gateway on the same state directory, and the message names the PID and port. The Gateway owns that state directory, so the CLI stops you rather than run a second writer against the same session store. Drop --local to go through the Gateway, or run openclaw gateway stop --force first if you really want the embedded runtime.
Cannot admit memory.cli for memory stores
openclaw memory status printed This operation requires exclusive offline state ownership while the Gateway was up, and openclaw mcp set failed the same way with Cannot admit mcp.set for MCP credentials and configuration. Same cause as above: run openclaw gateway stop --force, rerun the exact command, then openclaw gateway start.
Multiple agents are configured, but the skills command has no explicit owner
This appears as soon as a second agent exists, because skills are installed per workspace and the CLI will not guess which one you meant. Pass --agent <id> to skills check and skills list, and add it to scripts now so they keep working when someone adds an agent later.
Detected unsettled top-level await
openclaw uninstall --all --dry-run and openclaw reset --scope config --dry-run both stopped at a Yes/No prompt and exited with code 13 when run from a script with no terminal to answer. Add --non-interactive --yes; reset also needs --scope. With those flags the uninstall dry run exited 0 and listed its plan: remove the gateway service, ~/.openclaw and ~/.openclaw/workspace, after recommending openclaw backup create first.
DeprecationWarning: Plugin SDK
Several commands print [DEP_PLUGIN_SDK] DeprecationWarning lines in 2026.11.3, as they did in 2026.11.2. They’re warnings from bundled plugins, not failures, and the text gives 2026-12-31 as the last compatibility day before the old SDK calls go in the first January 2027 stable release. Because they go to stderr on otherwise successful runs, a script that treats any stderr as failure will break, so check the exit code instead. If you maintain your own plugin, that date is your deadline to move to the async calls the warning names.
If you also run coding agents next to OpenClaw, the Claude Code cheat sheet, Codex CLI cheat sheet and OpenCode CLI commands follow the same tested <tool> cheat sheet layout, reference tables first and lab output below. For the local model side, the Ollama commands cheat sheet covers pulling and serving the models OpenClaw talks to, and the official OpenClaw CLI reference lists every flag for the commands not covered here.