The Lava Agent in Slack: One Agent, Two Surfaces
The Lava Agent in Slack is the same agent you use in Lava Desktop, reachable from a Slack DM or a channel mention. It runs on your own Lava account, with your connected tools, your memory, and your balance. Installing the Slack app grants nobody any access to Lava by itself, because each person's own login is the entire authorization.
That last part is the design decision worth understanding, and it is the reason a shared Slack workspace does not become a shared agent.
Key Takeaways
- The agent is yours, not the channel's. Every person who talks to it in Slack connects their own Lava account first, and the agent acts as that person.
- Installing the Slack app grants zero Lava access. There is deliberately no matching between a Slack workspace and a Lava organization.
- Actions that send, change, or delete pause for approval in the Slack thread, and only the person who owns the run can approve them.
- Bystanders in a channel see that an approval exists, not the private values inside it.
- Slack is a control surface, not a replacement for Desktop. Long, tool-heavy work still runs better where you can watch it.
What does it mean that the agent runs on your account?
When you message the Lava Agent in Slack, Slack is only the doorway. The request goes to your Lava profile: the apps you connected, the memory from your past conversations, the wallet your calls bill against. Ask it about your calendar and it reads your calendar, not a shared team one.
This sounds obvious until you compare it to how most Slack bots work. A typical bot is installed once by an admin, holds one set of credentials, and acts as itself for everyone. Ask it to check the CRM and it checks the CRM through whatever account the admin connected. Permissions belong to the bot.
The Lava Agent inverts that. The Slack installation carries no Lava permissions at all. The first time you message the agent, you link your own Lava account, and from then on the agent in Slack is the same one you have been using at your desk.
Why this matters more than it sounds
If the agent inherited workspace-level access, then anyone in your Slack could reach anyone's connected tools by asking nicely. Binding the agent to each person's own login removes that entire class of problem instead of trying to police it with rules.
How do you add it?
There are two ways, and they answer different questions.
Add just for you. Installs the app if it is not already there, links your Lava account, and drops you into a DM with the agent, ready to work. Good for trying it out.
Add to your workspace. Puts the app in Slack for everyone. Each person still connects their own Lava account the first time they message it. Nobody inherits anything from whoever clicked install.
Both live in Lava Desktop under Settings. The second option is not an admin granting access, it is an admin removing the friction of everyone installing the app themselves.
What happens when the agent wants to do something consequential?
It stops and asks.
Reading is free. Sending an email, posting a message, moving money, deleting a record: those pause and render an approval in the Slack thread with buttons. Nothing runs until someone taps approve.
The approval is bound to the person who owns the run. If you are watching a colleague's agent work in a channel, you can see that it paused and what kind of action it wants to take, but you cannot approve it for them, and you do not see the private argument values inside the call. You see the shape of what is about to happen. The owner sees the contents.
Use channels for visibility, DMs for work
A channel is a good place to let a team watch an agent handle a recurring task, because everyone sees the pauses and the outcome. For anything with private inputs, a DM keeps the details where they belong.
Slack or Desktop: which should you use?
They are the same agent, so this is a question about the surface, not the capability.
| Surface | Runs as you | Approvals | See the work happen | Best for |
|---|---|---|---|---|
| Lava Desktop | ✓ | ✓ | ✓ | Long tasks, anything across app screens |
| Lava Agent in Slack | ✓ | ✓ | ✕ | Quick asks, approvals, team visibility |
The honest gap is the third column. In Desktop, the agent works in live signed-in tabs of your actual apps, and you watch it move. You can stop it, take the keyboard, and finish something yourself. Slack gives you a thread: the request, the pauses, the answer. That is enough for most short asks and not enough when you want to supervise closely.
There is a second gap worth naming. Work started from Slack runs in the cloud, and a cloud tool call carries more round-trip latency than the same call running on your own machine. For a three-step task nobody notices. For a thirty-step task, Desktop finishes sooner.
What does it actually look like in use?
The pattern we keep landing on internally is that Slack is where you start things and approve things, and Desktop is where you watch things.
Someone asks the agent in a DM to pull last week's signups and put them somewhere useful. It works, it pauses once before writing anything, they tap approve from their phone. Nobody opened a laptop.
The opposite case: reconciling a month of transactions across three tools. That belongs in Desktop, where the agent can move through the actual interfaces and you can see every step.
Character limit per Slack message chunk
Long answers split at paragraph boundaries rather than being dropped
That number exists because of a real failure. A correct, already-billed answer came back at 4,885 characters, Slack rejected it as too long, and the response was lost. Now long answers split cleanly at paragraph breaks, code blocks close and reopen across the split, and a delivery failure surfaces instead of vanishing. It is a small thing that only matters the first time it costs you an answer you paid for.
The bottom line
The Lava Agent in Slack works because it refuses to be a workspace bot. It is your agent, reachable from Slack, running on your account with your tools and pausing for your approval. Installing it grants nothing on its own, which means adding it to a shared workspace is a low-stakes decision rather than a security review. Use Slack to start work and approve it, and use Lava Desktop when you want to watch the work happen.
How Lava Helps
Lava Gateway is what lets one agent reach your tools and models from either surface, with every request metered and attributed to the person who made it. Because the agent runs on your account rather than a shared bot identity, usage and cost land where they belong instead of pooling into one anonymous total.
Lava Monetize handles the wallet side: balances, spending limits, and usage visibility per person. If you are rolling the agent out to a team in Slack, that is how you keep an eye on what it costs without turning every request into a budget conversation.
If you want the background on how an agent reaches your tools in the first place, what a gateway is explains the layer underneath both surfaces, and Introducing Lava Desktop covers the one where you watch the work happen.
