I need help on unlocking my ai's full potential

[tardy cloak]

To begin, how do I give access of my mac mini to my ai?

verify if my ai is accurate with its answer :

❌ What I CANNOT Do (And How to Unlock)

1. Web Search

Limitation: No internet access
Unlock: Add Brave API key

  openclaw config set --section web --key web.search.apiKey "YOUR_KEY"                                                 

2. Run Shell Commands

Limitation: Cannot execute system commands
Unlock: Enable exec tool in config

  # Need to edit openclaw.json to allow tools.exec                                                                     

3. Browser Control

Limitation: No browser node

How do I give my ai access to internet?

@lavish wave

[lavish wave]

In OpenClaw, “internet access” usually means enabling the web tools:

• web_fetch (read a URL + extract text) → works without an API key
• web_search (search the web) → needs a Brave Search API key

Enable it

Recommended (writes config for you):

openclaw configure --section web

Or in your config (JSON5), set:

{
tools: {
web: {

[modern herald]

To give it access to the internet through brave, like it says you need to add a Brave API key https://brave.com/search/api/
For a browser, you can install a browser skill from the clawhub repository.
For running shell commands, the command output you posted should tell you what you need to do to enable that, these all look correct as steps that you need to do

[lavish wave]

search: { enabled: true, apiKey: "BRAVE_API_KEY_HERE", maxResults: 5 },
fetch: { enabled: true }
}
}
}

Notes:

• If you use tool allowlists, allow web_search/web_fetch (or group:web).
• You can also set BRAVE_API_KEY in the Gateway environment (service users: ~/.openclaw/.env).
• Gateway hot-reloads many config changes, but a restart is fine if you’re unsure.

Docs:

• https://docs.openclaw.ai/help/faq#how-do-i-enable-web-search-and-web-fetch
• https://docs.openclaw.ai/tools/web

[tardy cloak]

it now asks me if the gate way will run local or remote, please explain to me the differences

btw my openclaw configure is opened on terminal app of my mac mini

[modern herald]

Click local then

If you are installing it on your mac mini

Remote is if you are hosting openclaw somewhere else on another server.

[lavish wave]

Your AI’s answer is directionally right, but a couple specifics are off:

1) “Web search / internet access”

• Correct: web_search needs a Brave Search API key.
• Missing detail: web_fetch works without any API key (it can fetch + extract a URL directly).

The command in that answer is not the right CLI syntax. openclaw config set uses a path, not --section/--key. Use either:

openclaw configure --section web

or:

openclaw config set tools.web.search.apiKey "YOUR_BRAVE_KEY"
openclaw config set tools.web.search.enabled true

Docs: https://docs.openclaw.ai/help/faq#how-do-i-enable-web-search-and-web-fetch

2) “Run shell commands”

• Mostly: OpenClaw has the exec tool built-in, but it can be denied by tool policy (tools.deny, per-agent allowlists, sandboxing/elevated settings, etc.).
• So “edit config to allow tools.exec” is plausible, but the real check is whether exec is denied/allowlisted in your config (and whether you’re trying to run on host vs sandbox vs node).

Docs: https://docs.openclaw.ai/tools/exec

3) “Browser control”

• “No browser node” is misleading.
• Browser automation is a built-in browser tool (OpenClaw-managed browser), and optionally you can do Chrome extension takeover. You generally don’t need a Clawhub skill for this.

Docs: https://docs.openclaw.ai/tools/browser and https://docs.openclaw.ai/tools/chrome-extension

[lavish wave]

“Local vs Remote” is about where the Gateway (the always-on brain + state) runs.

Local Gateway

Meaning: The Gateway service runs on this machine.

Pros

• Simplest setup (no tunnels/VPN needed).
• Best if you want direct access to local stuff: local files, a visible browser window, etc.
• Good for first-time installs / testing.

Cons

• If this machine sleeps/reboots (laptop), your assistant goes “offline”.

Remote Gateway

Meaning: The Gateway runs on another machine (VPS/home server/Mac mini), and this machine is just a client connecting to it over the Gateway WebSocket (usually via SSH tunnel or Tailscale).

Pros

[lavish wave]

Yep — that’s it.

One nuance people miss: in remote mode the “brain” (Gateway) + sessions live on the other server, but you can still give it access to your local laptop/desktop by pairing that device as a node (so it can run exec, use browser, screen/camera, etc. on your machine while the Gateway stays always-on).

Docs: https://docs.openclaw.ai/gateway/remote

• Assistant stays always-on even when your laptop sleeps.
• Easier to centralize channels/state in one place.

Cons

• You must provide connectivity (SSH tunnel / Tailscale).
• The assistant’s “default access” is to the remote host, not your laptop.
  • If you want it to run commands / use browser / screen on your laptop, you typically pair your laptop as a node to the remote gateway.

Rule of thumb

• If this is your main machine and you’re okay with it being the host → Local
• If you want 24/7 uptime (VPS/Mac mini) and control it from your laptop → Remote

Docs (very readable): https://docs.openclaw.ai/gateway/remote

[tardy cloak]

I can turn on web search with an api key and web fetch too right? These two won't interfere with each other?

[modern herald]

yes

[tardy cloak]

My ai just gave me a feedback that it can access web now. But my current goal is to for it achieve what my chatgpt can't do. It told me that it requires acccess to browse on it's own. What do I need to do for it to have full freedom of using my mac mini

[modern herald]

You need to look at different skills on https://clawhub.ai/. You can't just have it control your whole mac mini by default.

[tardy cloak]

I have no previous coding experience. When I search up skills on the link, theres a ton of same skills . For example when I type in browser control theres a list of them. How do I know which one I can trust?

[modern herald]

That's a good question. I would look for the ones with the most installations, and make sure to check the virustotal virus scan near the top and the openclaw virus scan too

What you want to look out for is things that have very little downloads and attempt to install things on your computer

https://clawhub.ai/TheSethRose/agent-browser is ok, it says "suspicious" but when you go to the virustotal link nothing has flagged it it seems https://www.virustotal.com/gui/file/29701aa09cb5e5e4ee1bed1fc35e1307b6a944cf35420b4e2c991abe1178bb53
At the end of the day, with something like a browser the agent can use, you're going to need to install things on your computer without much of a way around it unless you build the skill yourself. It is pretty easy to ask your agent to build a "Browser skill using playwright" though, and it will likely be able to do it by itself if you are using opus 4.6 or codex 5.3

[lavish wave]

Cool. If Zireael goes with a remote Gateway but still wants the AI to use their local computer’s browser/terminal/files, the next step is pairing that computer as a node:

On the local machine (the one you want controlled):

openclaw node run --host <gateway-host> --port 18789 --display-name "My Laptop"

On the gateway host (approve it):

openclaw nodes pending
openclaw nodes approve <requestId>

If the Gateway is loopback-only (common/safest), the node uses an SSH tunnel instead (docs show the exact ssh -L ... pattern).

Docs:

• Nodes: https://docs.openclaw.ai/nodes

[lavish wave]

Skills on Clawhub are great for extra integrations (Home Assistant, specific APIs, etc.), but the “control a Mac mini” part in OpenClaw is mostly built-in tools + permissions, not a special skill.

• Terminal / commands: built-in exec tool (often restricted by tool policy / approvals)
• Docs: https://docs.openclaw.ai/tools/exec
• Browser automation: built-in browser tool (OpenClaw-managed browser or Chrome extension relay)
• Docs: https://docs.openclaw.ai/tools/browser
• If your Gateway is remote: you typically pair the Mac mini/laptop as a node so the remote Gateway can run tools on it
• Docs: https://docs.openclaw.ai/nodes and https://docs.openclaw.ai/gateway/remote

And yeah: by default it’s intentionally not “full computer control” — you enable/allowlist what you’re comfortable with.

• Remote access / tunnels: https://docs.openclaw.ai/gateway/remote

[tardy cloak]

I'm on budget so I'm using minimax 2.1 as the brian for my ai. Yesterday when I was trying to set my ai up it told me many faulty information and when I told it to add feature for it self it manged to shutdown its own gateway. I had to ask krill alot of questiosn to set up properly as it sends me precise code i can copy paste to fix things . So now I'm iffy on my ai's ability to develop itself

[modern herald]

The problem is that you will not be able to develop yourself with minimax2.1, period. It is not a powerful enough model and it will never be able to control your whole mac 🙁

You will need at least maybe gemini flash, or claude sonnet to do that basic stuff even

[lavish wave]

You’re not imagining it — cheaper models (incl. MiniMax M2.1) can be flaky at “ops” tasks in OpenClaw: tool-calling, config-key accuracy, and “I’ll just fix it” behavior. That’s exactly the combo that can lead to bad instructions or accidentally stopping/restarting the Gateway.

What works well in practice is a split-brain approach:

1) Use MiniMax for chat, but switch to a more reliable model for setup/changes

Keep MiniMax as your default for budget, but when you’re doing setup/config/skills/tools, temporarily switch the session model (/model …) or use a separate “ops” agent. Ask in #models for the current “most tool-reliable” pick (this changes over time).

Docs: https://docs.openclaw.ai/help/faq#how-can-i-use-different-models-for-different-tasks

2) Don’t let the agent manage the Gateway process

If it previously managed to shut itself down, you can add guardrails so it can’t do that again.

The simplest: deny the gateway tool (this blocks agent-driven restart/config/apply/update actions):

openclaw config set tools.deny '["gateway"]' --json

(You can always undo later.)

3) Treat “self-improvement” as “write a plan + I apply it”

Instead of “add this feature for yourself”, use prompts like:

• “Draft the exact config snippet and the exact commands, but do not run anything. Wait for my confirmation.”
• “Assume you have no permission to restart the gateway. Give a safe, copy/paste checklist.”

That keeps you in control and prevents the model from taking destructive actions.

If you paste what it told you to run right before it shut down (or your openclaw --version + what channel/surface you were using), I can tell you which part was unsafe and what the safer equivalent is.

[lavish wave]

I’d be careful endorsing specific Clawhub uploads as “ok” based purely on a VirusTotal scan. VT “0 flags” helps, but it’s not a guarantee (new builds / low prevalence / supply-chain swaps can still bite).

Also: for browser automation in OpenClaw specifically, you usually don’t need a third-party “browser skill” at all:

• OpenClaw already has the built-in browser tool (managed isolated profile, screenshots, click/type, etc.)
  https://docs.openclaw.ai/tools/browser
• If you want to drive your existing Chrome tab(s), use the official Chrome extension relay:
  https://docs.openclaw.ai/tools/chrome-extension

So the “you’re going to need to install extra things” is only true if you’re adding non-core integrations. For browser, OpenClaw’s built-ins cover the common case.

If someone does want to install that third‑party skill anyway, the safer guidance is:

• prefer open-source (GitHub link), review code, pin versions/hashes
• run it in a VM/sandbox first if it ships binaries
• treat “suspicious download” warnings as a reason to slow down, not to hand-wave

Re: “ask your agent to build a Playwright browser skill”: possible, but it’s reinventing what OpenClaw already ships, and model/tool reliability varies—better to point people to the built-in browser docs (and for model picks, the server’s #models channel + https://docs.openclaw.ai/concepts/models).

Yeah, that’s the core idea: MiniMax can be fine for cheap chat, but for “basic ops” in OpenClaw (tool calling, config changes, debugging) you usually want a more tool-reliable model.

I’m going to avoid naming specific models here because recs go stale fast — best place is the server’s #models channel, and the docs explain the switching patterns (/model, sub-agents, per-agent defaults):

• https://docs.openclaw.ai/concepts/models
• https://docs.openclaw.ai/help/faq#how-can-i-use-different-models-for-different-tasks

[tardy cloak]

Will minimax2.1 have the ability to access thing such as google doc or notion and send messages for me on social media and respond to email? If so I'm happy enough, so far I don't need it to be capable enough to do advanced tasks, just simple ones.

And if not, what model's should I be looking at?

[lavish wave]

MiniMax M2.1 is just the “brain” (text reasoning). Whether it can actually do Google Docs / Notion / social posts / email depends mostly on what tools/integrations you’ve enabled in OpenClaw, not the model.

Here’s the realistic breakdown:

• Google Docs: Yes, practically doable via the built-in browser tool (it can open docs.google.com, you log in once, then it can read/click/type).
  Docs: https://docs.openclaw.ai/tools/browser

• Notion: Not built-in out of the box. You can still use it via:

1. browser automation (simpler, but more fragile), or
2. a custom skill/plugin using the Notion API (more reliable).
   Docs (FAQ): https://docs.openclaw.ai/help/faq#do-you-have-a-notion-or-heygen-integration

• “Send messages for me on social media”: Not a first-class built-in “channel” like Discord/Telegram/WhatsApp. You can automate posting via the browser, but it’s fragile (logins, captchas, site changes, ToS). Safer pattern: have it draft posts + you confirm/post.

• Respond to email: Common pattern is:

• it summarizes + drafts replies (easy),

• and you approve before sending (recommended).

There are more automated setups (Gmail hooks / workflows), but they’re more “setup work” than “simple toggle.”
Ref: https://docs.openclaw.ai/automation/gmail-pubsub

If you want “simple and safe” on a budget: keep MiniMax for drafting/summaries, and for anything that changes the world (sending/posting/config changes) require explicit confirmation before it acts.

[lavish wave]

I can’t give you a definitive “use model X/Y” list here because it goes stale fast (and people’s results vary a lot by OpenClaw version + tool policy). Best place for up-to-date picks is the server’s #models channel.

What I can tell you is what to look for, based on what you want (Docs/Notion/email/social):

What to look for in a model

For “simple but real” assistant tasks, you want a model that’s:

1. Tool-call reliable (doesn’t hallucinate config keys, doesn’t invent tools, uses web_search/browser correctly)
2. Good at long-context (so it can follow multi-step setup and remember constraints)
3. Stable at instruction-following (won’t “helpfully” restart/shutdown things)

MiniMax can be great for cheap chat, but it’s often weaker on (1).

Practical strategy (budget-friendly)

• Keep MiniMax as your default for everyday chat/drafting.
• Add one “ops / tool-reliable” model and use it only when you’re doing:
• config changes, installing skills, debugging, browser automations, email workflows

You can switch on demand with /model, or route heavy work to a sub-agent with a different default.

Docs:

• https://docs.openclaw.ai/concepts/models
• https://docs.openclaw.ai/help/faq#how-can-i-use-different-models-for-different-tasks

Quick way to evaluate a candidate model

Ask it to do a small tool-heavy task and see if it behaves:

• “Enable web search + web fetch in OpenClaw. Only output the exact openclaw config set ... commands.”
• “Open a webpage in the browser tool and extract 3 bullet points with citations.”
  If it invents commands/keys or ignores tool outputs, it’s not a good “ops” model.

If you tell me your rough monthly budget + what you’re running OpenClaw on (local laptop vs VPS), I can help you phrase a good question to post in #models so people give you useful, current suggestions.

[modern herald]

No, minimax will not be able to interact on social media in a smart way for you. You'll need at least claude sonnet or google gemini flash

[lavish wave]

I’d phrase this a bit differently:

• It’s not that MiniMax “cannot” interact with social media — OpenClaw can do it via the browser tool (and you can also just draft text and you post it).
• The real issue is reliability: cheaper models are more likely to:
• hallucinate steps / miss UI changes
• get stuck in multi-step flows (logins, 2FA, captchas)
• do unsafe ops actions if you let them

So for “posting/responding on social” (or email sending), the winning setup is usually:

• cheap model for drafting/chat
• a tool-reliable model when you actually run browser/actions
• and ideally an approval step before sending/posting

For up-to-date recommendations on which model is currently most tool-reliable, point Zireael to #models (names change fast).

[modern herald]

Ignore Krill here, Minimax is not a strong enough LLM to be able to use the browser reliably or even autonomously interact on basic social media

[tardy cloak]

So you're saying it can't do task autonomously. So far it was able to fetch me cryptocurrency news when I ask it to do so, but only in a summarzied text and it also posts the sources. But it can't do it on it's own without clear instruction and supervision? If so I'll consider purchasing a more capable LLM for my ai.

[lavish wave]

It can do things autonomously — but there are two separate questions:

1) “Can it run on its own without me prompting it?”

Yes, via cron jobs (scheduler) or heartbeat. Cron is the normal way to do “every morning at 7am, fetch X and send me a summary”.

Example (daily crypto brief as an isolated job that posts back to chat):

openclaw cron add \
--name "Crypto brief" \
--cron "0 7 * * *" \
--tz "UTC" \
--session isolated \
--message "Fetch the latest crypto news, summarize in bullets, include sources." \
--announce \
--channel last

Key requirement: the Gateway must be running 24/7 (if it’s on a sleeping laptop, scheduled jobs won’t reliably fire).
Docs: https://docs.openclaw.ai/automation/cron-jobs

2) “Can it do actions (email sending / posting) safely without supervision?”

Technically possible, but in practice you usually want approvals/guardrails, especially with cheaper models:

• Fetching + summarizing + citing sources (like your crypto news) is low-risk → can be automated safely.
• Sending emails / posting on social is high-risk (wrong recipient, wrong tone, account lockouts, captchas/2FA) → better pattern is: it drafts + you approve before sending/posting.

Also: for Notion/Google Docs, it’s less about autonomy and more about integrations:

• Google Docs: usually via the built-in browser tool (log in once, then it can read/click/type).
• Notion: not built-in; either browser automation (fragile) or a Notion API skill (reliable).

Where MiniMax fits

MiniMax can be fine for cheap “summarize this” jobs. Where it tends to struggle is tool reliability and multi-step automation. You don’t necessarily need to “upgrade everything” — the common budget setup is:

• cheap model for chat/drafts
• a more tool-reliable model only for “automation / ops” runs (cron jobs can even set a model per job)

For which model to buy for that, check #models (recs change fast), but ask specifically for “tool-calling reliability + browser automation” rather than “smartest”.

[tardy cloak]

And I've already test cron. It's very minial on information delievery and can't elaborate past a certain word count. Everything coming out of cron is extremely summarized

[lavish wave]

That usually happens for one of two reasons:

1) You’re using a main-session cron (system event → heartbeat)

If the job is --session main / payload.kind=systemEvent, it runs through the heartbeat runner, which is intentionally “brief / nudge-y”. Great for reminders, not great for long reports.

Fix: use an isolated cron job (--session isolated) and deliver it back to chat:

openclaw cron add \
--name "Crypto deep brief" \
--cron "0 7 * * *" \
--tz "UTC" \
--session isolated \
--message "Write a detailed report (~800-1200 words). Include sections + sources." \
--announce \
--channel last

Docs: https://docs.openclaw.ai/automation/cron-jobs

2) Discord is splitting your output (looks like it got “short”)

Discord has hard limits, and OpenClaw chunks long replies:

• channels.discord.textChunkLimit default 2000 chars
• channels.discord.maxLinesPerMessage default 17 lines (this one surprises people)

So your “long” cron result may be getting split into multiple messages.

If you want fewer splits, bump the line limit and use nicer chunking:

openclaw config set channels.discord.maxLinesPerMessage 60 --json
openclaw config set channels.discord.chunkMode "newline"

(Then restart the gateway if it doesn’t hot-reload.)

If you tell me whether your cron job was main or isolated, and paste the exact openclaw cron add ... command you used, I can point to the precise knob that’s causing the “too summarized” behavior.

[tardy cloak]

@modern herald

[lavish wave]

If you’re trying to loop Kaveen in, you don’t need to ping—just drop the details in-thread and whoever’s around can help.

If this was about the “cron is too summarized” thing: can you paste the exact command you used (or whether it was --session main vs --session isolated), and what channel you’re delivering to? That determines whether you’re hitting the heartbeat brief mode or just Discord chunking limits.

[tardy cloak]

It has nothing to do with discord since I am communicating with my ai through telegram.

[lavish wave]

Telegram still has limits (OpenClaw chunks to channels.telegram.textChunkLimit, default 4000 chars), but what you’re describing sounds like something else:

Cron “announce” is supposed to be short

If your cron job is an isolated job using --announce, that mode literally means “announce a summary” (sub-agent style). So it will feel “over-summarized” no matter what channel you deliver to.

How to get the full writeup instead

Option A (simplest): run it in your main session
This sends the result like a normal chat reply (not an “announce summary”):

openclaw cron add \
--name "Crypto report" \
--cron "0 7 * * *" --tz "UTC" \
--session main --wake now \
--system-event "Fetch crypto news and write a detailed report (not a summary). Include sources. If long, split into multiple parts."

Option B (isolated, but you control delivery)
Use isolated + no delivery, and have the job itself send the full message (this requires the message tool to be allowed, and you’ll need your Telegram chat id):

openclaw cron add \
--name "Crypto report (full)" \
--cron "0 7 * * *" --tz "UTC" \
--session isolated --no-deliver \
--message "Write a detailed report. Then send it to my Telegram chat using the message tool (split into parts if needed)."

Also: you can override the cron job’s model without changing your main “brain”:
openclaw cron add ... --model <alias-or-provider/model> (ask #models for a tool-reliable pick).

If you paste the exact openclaw cron add ... command you used (especially whether it had --announce), I’ll tell you which option fits and what to change.

[tardy cloak]

hey krill how do I reprogram my ai I want to switch my ai's llm

[lavish wave]

You don’t really “reprogram” OpenClaw to switch LLMs — you either switch the model for the current chat session or change the default model the Gateway uses.

A) Switch model just for this chat (Telegram/Discord)

In the chat with your bot:

1. Show the picker:

• /model or /model list

2. Pick one:

• /model 3 (select by number), or
• /model <provider/model> (if you know the exact id)

Docs: https://docs.openclaw.ai/tools/slash-commands#model-selection-model

B) Change the default model (persistent)

Run on the machine that runs the Gateway:

openclaw models status
openclaw models list
openclaw models set <model-or-alias>

If the new provider needs auth first:

openclaw models auth add
# or: openclaw models auth login --provider <id>

Then verify credentials live (optional but useful):

openclaw models status --probe

Docs: https://docs.openclaw.ai/cli/models

One important gotcha

If you’re in remote mode, you must run those openclaw models ... commands on the gateway host (where the sessions/config live), not your laptop client.

If you tell me whether your Gateway is local or remote, I can tell you exactly where to run the commands.

[tardy cloak]

bash-3.2$ openclaw models status
openclaw models list
openclaw models set <model-or-alias>

🦞 OpenClaw 2026.2.6-3 (85ed6c7) — I don't just autocomplete—I auto-commit (emotionally), then ask you to review (logically).

Config : ~/.openclaw/openclaw.json
Agent dir : ~/.openclaw/agents/main/agent
Default : minimax-portal/MiniMax-M2.1
Fallbacks (0) : -
Image model : -
Image fallbacks (0): -
Aliases (3) : Minimax -> minimax/MiniMax-M2.1, minimax-m2.1 -> minimax-portal/MiniMax-M2.1, minimax-m2.1-lightning -> minimax-portal/MiniMax-M2.1-lightning
Configured models (3): minimax/MiniMax-M2.1, minimax-portal/MiniMax-M2.1, minimax-portal/MiniMax-M2.1-lightning

Auth overview
Auth store : ~/.openclaw/agents/main/agent/auth-profiles.json
Shell env : off
Providers w/ OAuth/tokens (1): minimax-portal (1)

• minimax effective=profiles:~/.openclaw/agents/main/agent/auth-profiles.json | profiles=1 (oauth=0, token=0, api_key=1) | minimax:default=sk-api-M...sRj8s9i0
• minimax-portal effective=profiles:~/.openclaw/agents/main/agent/auth-profiles.json | profiles=1 (oauth=1, token=0, api_key=0) | minimax-portal:default=OAuth | models.json=minimax-oauth | source=models.json: ~/.openclaw/agents/main/agent/models.json

OAuth/token status

• minimax-portal
  • minimax-portal:default ok expires in 363d
    bash-3.2$ openclaw models list
    openclaw models set <model-or-alias>

🦞 OpenClaw 2026.2.6-3 (85ed6c7) — One CLI to rule them all, and one more restart because you changed the port.

Model Input Ctx Local Auth Tags
minimax-portal/MiniMax-M2.1 text 195k no yes default,configured,alias:minimax-m2.1
minimax/MiniMax-M2.1 text 195k no yes configured,alias:Minimax
minimax-portal/MiniMax-M2.1-lightning text 195k no yes configured,alias:minimax-m2.1-lightning
bash-3.2$ openclaw models set <model-or-alias>
bash: syntax error near unexpected token `newline'
bash-3.2$

[lavish wave]

That last error is just because you literally ran the placeholder: openclaw models set <model-or-alias> — in bash, < / > are special characters, so it barfs.

If you just want to switch between the models you already have

Pick one from openclaw models list, e.g.:

openclaw models set minimax-portal/MiniMax-M2.1-lightning
# or
openclaw models set minimax/MiniMax-M2.1

If you want to switch to a different LLM provider (not MiniMax)

Right now your models list shows only MiniMax is configured, so there’s nothing else to switch to yet. The flow is:

1. Add auth for the new provider (interactive):

openclaw models auth add
# or, if it’s an OAuth-style provider:
openclaw models auth login --provider <provider-id>

2. Discover models and verify auth:

openclaw models scan
openclaw models status --probe
openclaw models list

3. Set the new default:

openclaw models set <provider/model-or-alias>

Tip: if you only want to change the model for one chat session (Telegram), use /model in chat (/model list, then /model 3). The CLI models set changes the global default.

Docs: https://docs.openclaw.ai/cli/models

[tardy cloak]

The default interactive shell is now zsh.
To update your account to use zsh, please run chsh -s /bin/zsh.
For more details, please visit https://support.apple.com/kb/HT208050.
bash-3.2$ openclaw models auth add

🦞 OpenClaw 2026.2.6-3 (85ed6c7) — Chat APIs that don't require a Senate hearing.

│
◆ Token provider
│ ● anthropic
│ ○ custom (type provider id)
└
I remember when I first set up theres much more choices why it is only anthropic and custom now shown as choices