Cheatsheet: CostGuard and where your data goes

The CostGuard mental model

A software safety net that watches your BYOK spend against a monthly cap. Warns when you cross a threshold, blocks new sends at 100% (default). The cap-enforcement layer is live at production launch; pre-release builds show the numbers but do not yet enforce.

The three settings (Settings, Budget)

Setting	Options	Default
Monthly cap	$0 (unlimited), $5, $10, $25, $50, $100, custom	Pick what feels comfortable
Warn threshold	50%, 80%, 90%, off	80%
Hard-stop behavior	Block new messages, Warn-and-allow	Block

What counts toward the cap

Usage type	Counts?	Why
Pay-as-you-go API keys (Anthropic, OpenAI, Google AI, Groq, most others)	Yes	Per-token billing on your provider account; CostGuard sums it
OAuth (Codex via your ChatGPT subscription)	No	Your ChatGPT subscription pays; shows $0 on the Usage dashboard
Local models (running on your computer)	No	No provider in the loop; no dollar cost to track

You can mix providers freely. The cap looks at the paid ones only.

The two visual cues

State	Indicator	Behavior
Under warn threshold	Dock row indicator neutral	Send freely
At warn threshold (80% by default)	Dock row indicator yellow + one-time notification	Send still works; quiet visual signal
At 100% cap	Dock row indicator red	With Block on (default): next send rejected, banner appears in chat. With Warn-and-allow: every send warns but goes through

The hard-cap banner reads:

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

Unblocking once you hit the cap

Path	What it does
Raise the cap in Settings, Budget	Unblocks immediately on save
Wait for the period to reset	Default rolling 30-day window: oldest day’s spend ages out today. Alternative: calendar-month resets (matches monthly subscription billing)

Habits that compound to lower cost

Habit	Why it works
Pick the right model for the task	Heavyweight models cost 5x to 10x what smaller ones do per token
Fresh conversation when topic changes	Long threads carry full history with every new message; cost grows linearly with length
Use OAuth (Codex) where you can	Codex calls show $0 against the cap
Watch the Usage dashboard weekly	One specific session sometimes ate everything; that tells you what to change

One thing to know about the OAuth path: it 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.

The data path (one message, end to end)

your keyboard
     v
the Clawless app on your computer
     v
OpenClaw engine (still on your computer)
     v
the AI provider over the internet  <-- the only server in the path
     v
OpenClaw
     v
Clawless
     v
your screen

No Clawless server in the data path. Path is: your computer, the AI provider, your computer.

The three places your stuff lives

Place	What lives there	Who can read it
Your computer	API keys (encrypted by Clawless in OS secure storage); conversations, memories, agent definitions, settings (in the Clawless data folder, unencrypted; at-rest safety is the safety of the device, so turn on FileVault on macOS / BitLocker on Windows / LUKS on Linux if at-rest encryption matters)	You, plus anyone with physical access to the machine
AI provider servers	Your message and the memories that travel with it, for as long as it takes to generate a reply	The provider, per their terms; most default to not training on API customer data, but read the terms
Our license server	License state (trial, active, expired)	Us, for license validation only; no message content reaches here

The four-party trust model

You trust	With what
Us	The desktop app does what it claims; local data protected to the standards your operating system supports
The AI provider you picked	The contents of your messages and the memories that travel with them
OpenClaw (open-source engine)	Being a faithful intermediary; code is auditable, audits exist
Any tool, skill, or integration you install	The permissions it asks for, same model as browser extensions

Closed vs open: the distinction

Clawless is closed-source. The desktop app, the chrome around the engine.
OpenClaw is open-source. The engine bundled inside Clawless that talks to providers.

They are not the same project. The data path crosses both. The audit story is different for each.

If Clawless went away tomorrow

What	Status
Your conversations, memories, settings, agent definitions, API keys	Still on your computer. They do not vanish when the company does.
New app updates	Stop.
License check	Grace period built in for short outages. Long term: app eventually requires re-license.
Local models	Keep working offline, no cloud or license server needed. The strongest version of the no-cloud guarantee.

Pitfalls to dodge

Setting a cap so high it never trips (you learn nothing about your usage shape)
Forgetting OAuth (Codex) does not count toward the cap (your dashboard will be quieter than expected if most work is on it)
Treating Warn-and-allow as the safe default (Block is safer; Warn-and-allow is for specific reasons)
Comparing the Usage dashboard to the provider’s invoice down to the cent (close match expected; if there is a discrepancy, the provider’s bill is the legal truth, ours is a real-time tracking signal)
Assuming there is a cloud you can sign into from a second device (there is not; local-first trades cross-device sync for the no-cloud guarantee)

Worth opening once

The Usage dashboard, weekly for the first month. Look for the one session that ate disproportionate budget. That is the most important number on the screen.

What lands next

You have reached the end of the Getting Started track. The privacy track unpacks the “no Clawless server in the data path” model in full architectural detail. The AI literacy track (“AI Foundations”) covers what models actually are and how they work. Lesson 3 of Track 22 (“Building with Claude”) goes deeper on cost-and-capability tuning if cost optimization is now an interesting topic to you.