Product
I'm Mitchell, the founder at Lava. The short version: it's a desktop app that puts an agent next to the apps you're already logged in to, and it does the work through a direct connection to each app instea…
I'm Mitchell, the founder at Lava. The short version: it's a desktop app that puts an agent next to the apps you're already logged in to, and it does the work through a direct connection to each app instead of driving a browser. You describe a task, it reads your real data, builds a plan, and executes it. Scoped permissions: reads run freely, but writes, sends, deletes, and payments need your approval. A concrete one: you say "follow up with everyone from last week's demos" and it reads your CRM and your calendar, drafts the emails in Gmail, and posts the summary in Slack. That's one conversation, not four tabs. There are 200+ providers connected today, covering all the apps your team already uses, so the same shape works for whatever your stack actually is. Why a direct connection and not browser automation: we started with a browser agent and it was miserable. Every time a site updated its layout, something broke. It was slow, and there was no clean way to tell the user what was about to happen. A direct connection to each app means every step is a specific action with known inputs, which is what makes the approval screen possible at all. The tradeoff: we only support what an app makes available through its connection, and we had to build a connector per app instead of getting everything for free. What makes this different from the alternatives: you don't need an engineering team to set up a workflow. You describe what you want in plain language and the agent plans and executes it. You can switch between AI models, so you're not locked into one provider. For apps that don't have a connector yet, Lava can also work through your open browser tabs, so the agent isn't limited to what has a connector. And any conversation can become a scheduled routine that runs on its own. To be clear: the download is free, but the agent runs on credits or a subscription. You need an account for the agent to work. The part I'd most like feedback on is the permission model. Permission fatigue is a real problem and there's been a lot of discussion on HN about it. We use scoped permissions with three modes: always ask, ask before sending (the default: reads run freely, anything that writes, sends, deletes or pays stops and asks), and never ask. The default is the safe one and also the annoying one. The approval gate is also what stops a prompt-injected agent from doing something you didn't ask for. If every write needs your sign-off, the attack surface shrinks. Scheduled routines run without a prompt because nobody is there to answer, and I'm not fully comfortable with that asymmetry. If you've shipped something in this shape, I'd like to hear where you landed. Practical notes: direct download for macOS and Windows. The build is signed and notarized. Happy to trade notes with anyone building agents that touch real user accounts.
Discuss this product