Lava: An AI Agent Workspace for the Apps Your Team Already Trusts
The Workspace is live, and it is built for the way your operations team works
An AI agent workspace is a single place where an agent can see your team's tools, complete multi-step work across them, follow the approval mode you choose, and run routines in the background. The Lava Workspace is that: it replaces the tab juggling, not the tools themselves.
The workspace itself is a desktop app for Mac and Windows, and the agent works directly inside the tools your operations team already signs into: Gmail, Slack, Google Sheets, Notion, Linear, GitHub, and your CRM. You describe the work in plain language. The agent opens the right tabs, reads the live state, finishes the steps, and follows your chosen approval mode. If your team spends its week on recaps, follow-up queues, and CRM hygiene across a stack of SaaS tools, this was built for the boring majority of the job, and the boring majority is where the hours actually go.
Key Takeaways
- Available for Mac and Windows. Install it once, sign in to the apps your team already uses, and the agent works inside them on your own accounts. No migration plan, no new dashboard to babysit.
- API or browser, whichever the tool has. Agent Connectors call tools that publish an API; for the rest, the agent works in the signed-in tab in front of it.
- Routines keep working even when your laptop is closed. Schedule the weekly recap once and it keeps running in Lava's cloud: the Thursday recap still lands while you are on the Friday flight.
- You keep the brakes. Choose Always asks, Asks on sends, or Never asks. The mode decides when the agent needs approval. Never asks lets it act without approval prompts.
- Model choice stays with you. Pick Claude, GPT, Gemini, or open-weight models per task, and swap them later without re-architecting anything.
What's new
The workspace ships as four connected pieces, and together they explain why it behaves differently from a chat window.
Lava Desktop for Mac and Windows
The desktop app is where the workspace lives. The agent sits beside the apps you have signed into, reads what each tab currently shows, and moves across them in sequence, so "the email from tomorrow's vendor call" means something concrete.
Agent Connectors, with a browser fallback
Connectors reach the tools that publish an API. For everything else, the agent acts inside the tab itself. That combination is why this works with a real ops stack, not just the handful of tools somebody built a polished integration for.
Routines that run with the laptop closed
Routines run on Lava's servers on a schedule of their own. Write the workflow once, set the cadence, and the run happens when it should. If a step cannot finish safely, the routine reports what stopped instead of improvising.
The same agent in Slack
The agent is reachable in Slack with the same profile and the same approvals. How it works, and where Slack is the better control room, is in The Lava Agent in Slack.
None of the four is a separate product, and you do not have to adopt them all at once.
The first test most ops people run
Pick the most tedious recurring task you own: the Friday end-of-week recap, or the Monday triage of the ticket queue. Set it up once as a routine, schedule it, and let it show up on time without you. If it does not prove itself in two weeks, delete the routine. That is the entire commitment.
What this changes for your operations team
Three reasons, in the order your team notices them.
- The agent comes to the work, because the work already lives in seven tabs. Your team does not need a new source of truth, and finance does not need to re-approve another vendor stack. Lava connects with your existing accounts, so the setup is the login you have today. There is no data migration, no export script to maintain, and no "lovely, now let's re-implement SSO" meeting at the end.
- The recurring work stops costing attention. Three concrete examples:
- The weekly ops recap. Pull usage numbers from your product data, ticket counts from the tracker, unresolved escalations from the inbox, the three metrics your CEO asked about last week, and have the draft in Slack before Friday's meeting instead of at 9:12 that morning.
- The follow-up queue. Point the agent at unanswered customer threads. It drafts a reply for each one and queues the drafts for review, so two hours of chasing becomes twenty minutes of approving.
- CRM hygiene. After a sales call, the agent adds your notes to the CRM and updates the deal's stage. With Always asks selected, you approve each change before it happens. The CRM stays the system of record; it just stops falling behind reality.
- The browser fallback keeps the long tail covered. Internal scripts break when a page moves, and a vendor without an API was previously a dead end. An agent working in the signed-in tab only needs your login, so "there is no API for that" stops being "there is no coverage for that."
Where the advantage actually compounds
If the recap eats two hours every Friday and the follow-up queue another two on Monday, that is four hours a week of mechanical work waiting for a routine. Hand it over once and the same people cover the judgment calls at the top of the stack, the ones nobody has time for today.
How control works: approvals, modes, and models
Approval requirements depend on the mode you choose.
Always asks pauses before each change. Asks on sends pauses before sends, deletions, payments, and external sharing. Never asks runs without approval prompts. Scheduled cloud routines also run unattended, without approval pauses.
For interactive tasks, you can choose the model. A workspace is not married to one vendor. Draft on Claude, summarize on GPT, work through planning on Gemini, and route to open-weight models where policy calls for it.
Does the workspace replace our stack?
No. It sits on top of your current stack and reads live state through your own logins. The systems of record stay exactly where they are. If a tool gets dropped next quarter, the routines that touch it change and the data does not move.
What does setup look like?
Install Lava Desktop on Mac or Windows, sign in, and start with one job that repeats. Connect your apps as the agent asks for them; it tells you when it needs one it cannot reach. If you want the category mapped out before you commit, how agent workspaces compare is the honest side-by-side.
Start with one routine, not five
The teams that get value fastest start with a single job they own end to end: the recap, the queue, or the handoff list. Prove one routine, then let it ask you what should run next week. The workspace earns the rest of its surface area one confirmed job at a time.
Your teammate. Your rules.
The whole argument fits in one sentence: your operations team does not need to migrate anything to get an agent. The workspace rides on the tools your team already trusts, follows your chosen approval mode, and keeps the recurring jobs moving while the laptop is closed.
We built it this way because operations work cannot stop while AI gets installed. The recaps still send, the queue still moves, the follow-ups go out on time, and the approvals sit where the decisions live, all on top of the stack your team already runs.
That is the trade Lava Workspace is built around: more gets done with the tools your team already uses, and the rules stay yours.