Lesson: CostGuard and where your data goes

Two anxieties tend to surface in the first week with Clawless. The first is “what if I accidentally rack up a giant bill?” The second is “who is actually seeing this conversation?”

Both are fair. Both have specific answers. This lesson covers them.

About twelve minutes to read. No code, no setup beyond what you already have.

The cost worry, and what CostGuard does about it

You read in lesson 3 that Clawless is BYOK: you bring your own API key, and the AI provider bills you directly for every message. That model is honest and it has no markup. It also means there is nothing standing between you and a real charge on your card if something goes wrong.

Most of the time nothing goes wrong. You type, the agent answers, the cost is fractions of a cent. But “most of the time” is not “always.” A long agent loop that decides to do twenty tool calls in a row, a runaway conversation that takes a wrong turn, an experimental prompt that produces a 10,000-word reply: any of these can be more expensive than you expected. Not catastrophic. But more than zero.

CostGuard is the safety net. You set a monthly spending cap, and Clawless makes sure your AI usage does not go past it. If you get close to your cap, you get a warning. If you hit it, new messages will be blocked until you raise the cap or the next month begins. It is a small piece of software whose entire job is preventing the kind of surprise that ruins a good day.

A quick note on timing. The cap-enforcement layer of CostGuard is being built in time for production launch. Pre-release builds of Clawless show you the usage numbers (so the picture is real) but do not yet enforce the cap automatically. If you are on a pre-release build, treat the cap as a personal discipline until the enforcement layer ships; the numbers are correct, the block-on-cap behavior will be live at launch.

Setting your budget

Open Settings from the gear icon at the bottom of the navigation rail on the far left. In the sticky sidebar on the left of the page, click the Budget section.

Three settings to pick.

Monthly cap. A dropdown with sensible options: $0 (unlimited, equivalent to no cap), $5, $10, $25, $50, $100, or a custom amount. Start somewhere low if you are nervous. You can raise the cap at any time, and raising it unblocks immediately if you have already hit a previous lower one.

Warn threshold. When to notify you that you are getting close. The options are 50%, 80% (default), 90%, or off. The 80% default works for most people: enough headroom to react, not so frequent that you tune it out.

Hard-stop behavior. What happens when you hit 100% of the cap. The default is “Block new messages”: the next send is rejected with a friendly banner saying you have reached your budget, and you have to raise the cap (or wait for the period to reset) to keep going. The alternative is “Warn and allow”: every send over the cap shows a warning, but the message still goes through. Pick this if you would rather be informed than stopped.

Save the three settings and CostGuard is active.

What counts toward the cap

Not every message you send racks up cost the same way. The cap follows the AI provider invoice rules.

Pay-as-you-go API keys (the standard mode for Anthropic, OpenAI, Google AI, Groq, and most other providers). The dollar cost of every message counts toward the cap.
OAuth providers, specifically Codex when you signed in with your ChatGPT account from lesson 3. Those sessions show $0 (OAuth) on the Usage dashboard because billing is handled by your ChatGPT subscription, not by per-API-call charges on a key. They do not count toward your CostGuard cap.
Local models (running on your computer with no cloud provider). Cost shows as $0 because there is no provider bill; these do not count.

You can mix providers freely in the same Clawless install. The cap looks at the paid ones only.

What happens at the warn line and the hard line

Two visual cues track your spend through the month.

When you cross the warn threshold (80% by default), a notification appears once. The dock row indicator turns yellow as a quiet visual signal that you are getting close. Nothing else changes. The yellow indicator stays yellow until you either raise the cap, drop below the threshold again as messages roll out of the period, or hit the hard line.

When you reach 100% of the cap, behavior depends on the hard-stop setting you picked.

With Block on (the default), the next message you send is rejected. A friendly banner appears in the chat:

Monthly budget reached. Bump your cap in Settings to keep going.

The dock row indicator goes red. Sends stay blocked until you raise the cap (which unblocks immediately) or the next reset comes (which unblocks at the new period).

With Warn-and-allow on, every send over the cap shows a warning, but messages still go through. Use this mode if you have a specific reason to allow a temporary overage and you would rather get the warnings than the wall.

How the period resets

The default is a rolling 30-day window. Every time you send a message, Clawless looks back 30 days and sums the cost; if the total is under your cap, the message goes through. This means there is no “reset day” you have to wait for; cost from 31 days ago aged out as of today.

If you would rather the budget reset on the first of every month (the calendar-month model that matches most monthly subscriptions), you can change to calendar-month resets in Settings, Budget. Both modes are reasonable; pick the one that matches how you mentally track recurring costs.

A few habits that keep costs down

A short list of small things that compound.

Pick the right model for the task. Heavyweight models cost five to ten times what smaller ones do per token. A quick factual question to a top-tier model is the same answer at a fraction of the cost on a smaller one. Lesson 2 walked the model picker; lesson 3 of Track 22 (“Choosing your model and the effort dial”) goes deeper on the per-task economics if you want them.

Start a fresh conversation when the topic changes. Long conversations carry the entire prior thread as context with every new message. The cost grows roughly linearly with conversation length. A long tangential thread costs more than two shorter focused ones.

Use OAuth where you can. If you have a ChatGPT subscription, the Codex path from lesson 3 makes those calls $0 against your CostGuard cap. Same models, no per-token charge. One thing to know: the OAuth path bills through your ChatGPT subscription, so it is governed by that subscription’s consumer terms rather than the API terms; worth a glance if your messages are sensitive.

Watch the Usage dashboard. The Usage panel on the navigation rail breaks down spend by provider, by model, and by conversation. After your first week with Clawless, opening it once tells you a lot. Sometimes the answer to “where did my budget go?” is “one specific session ate everything,” and that tells you exactly what to change.

Now the privacy half: where your data actually goes

The cost question is one half of the worry. The other half is “who sees this conversation?” The honest answer is short. It is also worth understanding clearly, because it is the most important thing to know about Clawless after you know how to send a message.

When you send a message in Clawless, the path is:

Your message goes from the chat input to the Clawless app on your computer.
The app passes it to an open-source engine called OpenClaw, which is bundled with Clawless.
OpenClaw sends the message to the AI provider you have selected, over the internet.
The provider’s model thinks, generates a reply, and streams it back.
OpenClaw passes the reply to Clawless.
Clawless renders the reply in the chat.

Nothing in that path goes through a Clawless server. The path is: your computer, the AI provider, your computer. There is no Clawless cloud holding your conversations.

Where your stuff actually lives

Three places matter, and it is worth being clear about all three.

Your computer. This is where most of your data lives. Your API keys, encrypted using your operating system’s secure storage. Your conversations. Your memories (from lesson 6). Your agent definitions. Your settings. All of it is on the disk in front of you. One thing to know: only the API keys are encrypted by Clawless; conversations, memories, and settings sit in the Clawless data folder unencrypted. Their safety is the safety of the device they live on. If at-rest encryption matters for you, turn on your operating system’s disk encryption (FileVault on macOS, BitLocker on Windows, LUKS on Linux).

The AI provider’s servers. Every time you send a message, the message and any context (your memories that travel with it, the prior conversation, any attachments) go to the provider’s servers. They process it on their side and stream back a reply. What the provider does with your messages after that depends on their terms; most providers default to not training on API customer data, but read their terms anyway, especially if your messages contain anything sensitive.

Our license server. This is the one piece of state we keep on a server we run. It is purely for license validation: knowing whether your install is on a trial, active subscription, or expired. No conversation content, no memories, no message bodies ever go there. They never leave your computer, so they never reach us.

What you are trusting to whom

A useful framing. When you use Clawless:

You trust us to ship a desktop app that does what it claims and protects your local data to the standards your operating system supports.
You trust the AI provider with the contents of your messages. They see what you type and what your memories carry along.
You trust OpenClaw (the open-source engine) to be a faithful intermediary. Because it is open source, you can audit the code yourself or read other people’s audits; if you do neither, you are trusting the maintainers and our review of their work.
You trust any tool, skill, or integration you install with whatever permissions it asks for. That is the same trust model you would apply to any browser extension or app plugin.

This is roughly the same trust model as any software that uses third-party AI APIs. The short version: pick providers you actually trust with the contents of your messages, and be cautious about installing tools and skills from sources you do not recognize.

What happens if Clawless shuts down

This is worth saying out loud because the question comes up often.

Your conversations, memories, and settings are on your computer. They are not in any Clawless cloud. If the company went away tomorrow, your data would still be there, in the same place it has always been, readable by you. The app would keep working for as long as your local install runs and the AI providers you use stay online. The license check has a grace period built into it, so a brief server outage on our end will not lock you out.

The strongest version of this guarantee is local models. Clawless supports running models entirely on your own computer with no cloud provider in the loop. You can launch and chat completely offline, which doubles as the answer to “what if the internet is down” and the answer to “what if I need a setup that does not depend on anyone else’s servers.”

Common first-day surprises

Five things people often notice in the first week.

OAuth (Codex) sessions show $0 on the Usage dashboard. This is correct, not a billing bug. OAuth means your ChatGPT subscription is paying for those messages; there is no per-API-call charge on a Clawless-tracked key. They also do not count toward CostGuard. If most of your work is on Codex, the dashboard will be quieter than you expect.
The dock row indicator changes color at the warn line, not at the hard line. Yellow at 80% (default), red at 100%. If you only ever look at the red signal, you will be reacting at the worst possible moment.
Raising the cap unblocks immediately. You do not have to “save up” to the new cap. The next send goes through the moment Settings is saved.
The numbers on the Usage dashboard come from Clawless’s own per-message tracking, not from the provider’s billing API. They should match the provider’s invoice closely, but if there is ever a discrepancy, the provider’s bill is the legal truth; ours is a tracking signal that helps you make decisions in real time.
Your conversations are on your computer, not in a cloud you can sign into from another device. If you install Clawless on a second computer, you do not see your phone’s conversation history (and there is no phone version). The local-first model trades cross-device sync for the no-cloud guarantee.

What you should remember

CostGuard is your spending safety net. Set a monthly cap, a warn threshold (default 80%), and a hard-stop behavior (default Block). Pay-as-you-go provider calls count toward the cap; OAuth (Codex) sessions and local models do not. The cap-enforcement layer is live at production launch; pre-release builds show usage but do not yet enforce the cap.
The dock row indicator tells you where you are. Yellow at the warn threshold, red at 100%. The banner at 100% reads “Monthly budget reached. Bump your cap in Settings to keep going.”
The data path is: your computer, the AI provider, your computer. There is no Clawless cloud holding your conversations. Three places matter: your computer (where most things live), the AI provider’s servers (where your message goes to be answered), and our license server (license state only, no message content).
If Clawless went away tomorrow, your data stays. Conversations, memories, and settings are on your computer. Local models let you keep working offline if you ever need to.

In the practice that goes with this lesson, you set a monthly cap that feels comfortable, open the Usage dashboard to see what your week looked like, and find the privacy policy from the in-app About panel. About twenty minutes start to finish.