Google shipped computer use in the Gemini Mac app on Thursday. Gemini can open apps, work through menus, fill in forms and move information between local files and websites. The same launch notes spell out the catch: a hostile document or website could talk the agent into acting against you. For anyone building agents that touch money, the interesting part is Google's list of controls, because it is also a checklist of what a wallet-holding agent needs. (Decrypt)
One note on scope. Decrypt's report does not mention crypto, wallets or digital assets in connection with this feature, and it does not say whether Gemini can reach wallet apps. What follows is our read on the pattern, not a claim about what Gemini does with a wallet.
What Google says it shipped
According to Decrypt's account of Google's announcement:
- ▸What it does: "Just describe what you need, and Gemini handles the clicking, typing, menus, and forms for you."
- ▸Where it runs: an "isolated background window," so the main display and cursor stay free. You get a live preview in the chat and can interrupt a task that drifts.
- ▸Who gets it: English, US, adults with personal paid Google accounts, on an Apple Silicon Mac running macOS Sequoia 15.0 or later. Work and school accounts are not supported at launch.
- ▸How you start: enable computer use in the Mac app, open Task mode, prompt with "@Computer use," and grant macOS Accessibility and Screen & System Audio Recording permissions.
Google's examples are deliberately mundane: pulling shipping details from PDFs into online forms, editing photo batches, building presentations from dashboard data.
The controls, and the gaps Google names itself
Google describes per-app permissions: Gemini asks before controlling a new app, and you can deny, allow once, or allow on an ongoing basis. There are allowed and blocked app lists, and Google says sensitive actions require confirmation.
Then the warnings, which come from Google's own documentation as reported by Decrypt:
- ▸The agent can reach active sessions in desktop applications.
- ▸It may encounter authentication codes displayed on screen.
- ▸Some actions cannot be reversed in certain apps.
- ▸Prompt injection: malicious instructions hidden in a document or website could persuade the AI to act against the user's wishes, and with access to several apps, content in one program could drive actions in another.
- ▸Optional on-demand file backups "cannot always be created." Where they work, the temporary copies last 24 hours or until the next prompt, and some apps may block backup of their content.
Decrypt also notes that an app set to "always allow" can be accessed again without another app-level prompt. And a background window is not, on its face, a guarantee that files or accounts are isolated from risk.
Credit where due: publishing that list is the honest move. Most agent launches bury it.
Why this matters for onchain agents
Swap "desktop apps" for "wallets and exchanges" and every line above gets sharper.
- ▸Active sessions are the whole ballgame. A logged-in browser wallet or exchange tab is an active session. An agent that inherits your session inherits your authority.
- ▸Irreversible is the default onchain. Google flags that some app actions cannot be undone. A signed transaction is that, by design. There is no 24-hour temporary copy.
- ▸Prompt injection is cross-app by nature. The risk Google describes, hostile content in one place steering actions in another, is the same shape as a poisoned web page steering a wallet-connected agent.
- ▸"Always allow" is a trust decision. Allow once, allow always, blocked lists: these are spend-limit-style ideas, applied to apps instead of dollars.
The take: a screen-driving agent is the most general interface an agent can have, and the least precise. It authorizes by *application*. A wallet agent should authorize by *action*: which asset, which counterparty, what ceiling, what expiry. "The agent may use the exchange app" is a much bigger sentence than "the agent may move up to X of this token to this address today."
If your agent can move funds, "trust me bro" is not a permission model, and neither is a pop-up it clicks through itself.
What a builder can take from this
If you are building an agent that holds or moves value, the checklist Google's launch implies looks like this:
1. Scope by action, not by app. Use wallet controls with explicit limits rather than handing the agent a session. 2. Keep keys out of the screen. Prefer a signer with its own policy over an agent that drives a UI where codes and sessions are visible. 3. Treat every input as untrusted. Documents, web pages and tool output can carry instructions. Separate reading from signing. 4. Make irreversible steps explicit. Require a human or policy confirmation before anything that cannot be undone, and log what was approved. 5. Preview before you commit. Google gives users a live preview and a stop button. A transaction simulation is the onchain version.
This is the part of the stack Sato maps: the wallets, skills, MCPs and payment rails that give an agent narrow, checkable authority. The prompt is the cute part. The permission layer is where the work lives. Describe the agent and see what it would take at satohub.ai/build.
What to watch
- ▸Whether work accounts and wider regions follow. Launch is US, English, personal accounts only. Broader rollout would put these controls in front of far more setups.
- ▸How "always allow" behaves in practice. Google's documentation concedes a remembered grant skips the app-level prompt. Whether sensitive-action confirmations hold up across apps is the part to test.
- ▸Operating-system-level limits. Per Decrypt, Apple introduced tighter access controls in the macOS 27.3 beta three days earlier, letting users restrict permissions to specific apps or folders and require fresh approval. Platform-level gates change what any agent can reach.
- ▸Desktop-control agents from other labs. Decrypt names OpenAI's Codex, Anthropic's Claude and Kimi Work as comparable desktop-control tools, so expect the permission questions to travel with them.
Self-reported safeguards are claims until someone tests them. Read the documentation, then read it again as an attacker would.