How Lava Desktop Works

Lava Desktop puts a live, signed-in copy of each of your apps in a sidebar, alongside the agent. When you give the agent a task, it works across those apps to get it done. This page explains the mechanics behind that: how the agent actually acts, why it stops and asks before high-stakes actions, and what costs money.

​

How the Agent Gets Work Done

The agent has two ways to interact with your apps: through the Lava gateway, and on the live page in front of it. Through the Lava gateway. Most real work happens here. When the agent sends an email, creates a calendar event, updates a contact, or takes any action that changes something, it calls that app’s API through the Lava gateway. The gateway authenticates the request, checks your balance, calls the app on your behalf, and settles the cost against your balance. This is how the agent can act reliably on your behalf, even in apps that don’t offer a click-based automation path. On the live page. The agent can also read the live tab to gather context: the content of an open email thread, the current state of a document, a contact record you have open. It uses what it sees on the page to ground its work, and it shows results back on that same tab when the task is done. Most tasks combine both: the agent reads context from the live page, then acts through the gateway to do the work.

​

The Path of One Task

Say you ask the agent to send a follow-up email based on your last conversation with someone. The agent reads the thread from the live tab, drafts the email, then pauses for your approval before it sends. After you approve, it routes the send through the gateway. Your balance is settled after the call completes.

​

Why High-Stakes Actions Pause

By default, in Normal mode, the agent works through a task on its own and pauses only before high-stakes actions: sending a message or email, deleting data, making a payment, or sharing something externally. Routine reads and low-risk writes (creating a draft, updating a record, adding a calendar event) run without a pause. The diagram above shows an email send, which is a high-stakes action. The agent pauses, shows you an Approve action card, and waits for your go-ahead before it sends. That pause is not a limitation of the agent. It is the designed behavior for actions that are outward-facing or hard to reverse. The approval card shows you exactly what the agent is about to do. You can approve it, decline it, or redirect the agent. If you approve, the action goes through the gateway. If you decline, the agent stops that step and waits for your next instruction. You can adjust how the agent pauses from the mode dropdown in the composer footer:

• Normal (default): pauses before sends, deletes, payments, and external shares only.
• Careful: pauses before every write, including drafts and record updates.
• Auto: runs the entire task without pausing, including sends, deletes, and payments.

For the mechanics of reviewing and approving specific actions in the moment, see Working with the Agent.

​

The Agent Never Acts on a Guess

When the agent needs to reference something specific (a person’s name, a record, a message thread), it reads the live page or asks the connected app through the gateway to find it. If it cannot ground the reference with confidence, it asks you before it proceeds. This means the agent will sometimes pause mid-task with a question like “I found two contacts named Alex. Which one did you mean?” That pause is intentional. Acting on a wrong assumption in your real accounts is worse than asking.

​

What Costs Money

Every task the agent runs draws from your balance in two ways:

• Model calls. The agent uses a language model to understand your request, plan the steps, and compose its responses. Each exchange draws a small amount from your balance.
• Provider calls. When the agent acts through the gateway (sending an email, querying a calendar, updating a CRM record), the app’s API call is billed from your balance.

Reads and drafts that stay within the app’s live page don’t add provider call costs, but model calls still apply. A larger task that calls several apps costs more than a quick lookup. Your balance goes down as work happens, not as a lump sum up front. For how to add funds, check your balance, and understand what your balance covers in more detail, see Funding Your Balance.

​

When This Model Matters

Once you have this picture, some behaviors that might seem odd become predictable:

• The agent switches the active tab during a task because it is reading context or showing you a result on the live page.
• A task stops partway through with a question because the agent hit a reference it couldn’t resolve from the page alone.
• Your balance drops a little after each task, not all at once, because each API call settles separately through the gateway.
• The agent moves quickly through reads and low-risk writes but pauses before sends, deletes, payments, and external shares, because those are the actions the gate applies to by default.

​

Next Steps