Openbase - Manage your team of AI agents by voice, from anywhere
by•
Openbase lets you manage a team of AI coding agents by voice, from anywhere; no screen, no desk.
AI agents can work async, but supervising them still pulls you back to a laptop. OpenBase fixes that. Dispatch tasks, steer agents mid-work, and approve changes just by talking. The agents write code, control your computer, open PRs, commit and push; you stay in control from your phone.
Works across providers and syncs to your machine, so you pick up at your desk right where you left off.
Replies
Openbase
Update: we've also just uploaded our YouTube demo to our launch, check it out here:
Creatium
Great idea, it’s the worst when you’re running an agent process and step away, screen goes dark and all work pauses. Best of luck!
Openbase
@kelly_king3 Thanks Kelly!
Dial
good, that's the right default. re-asking is a small friction cost that's obviously worth it compared to the alternative on something like a merge
Creatium
Congrats on the launch @lucaszhao! Is this mostly for checking in and approving things while you're away, or can you actually kick off real work from scratch without touching a laptop? I look forward to seeing where you take it.
Openbase
@gabo_deodato1 Thank you Gabo! Our idea here is you should be able to kick off agents to perform any task away from your laptop, at (hopefully) the same speed as if you were at your desk. So you are able to do both, just depends on your work and preferences.
Congrats on the beta! Getting to approve a PR out loud on a walk instead of hunting down a laptop is a great use of voice as an interface.
I'm curious what stops a sensitive approval from firing on someone else's voice, a recording, or an accidental phrase match before it reaches a real repo. That's the piece of the voice layer I'd want locked down first.
@vollos Great question. Today, you can set a private safety word that must be spoken before a sensitive action is approved, which helps prevent accidental phrase matches or someone nearby triggering it. That said, it isn’t the same as verifying your actual voice, so speaker verification and replay protection are still areas we want to strengthen.
@zoky_zhou The safety word closes the accidental-trigger case nicely. On the replay side you flagged, I'm curious what else gets checked when a sensitive approval comes in, like whether it also has to come from a paired device or an active session, versus the phrase alone being enough to authorize it right now. A captured recording or an overheard word matters a lot less if something else has to line up too.
@vollos Yes. Approvals have to come from your active, authenticated session on your device. The gap we're still closing is verifying it's actually you speaking within that session (voiceprint/replay protection).
@zoky_zhou Good to know real actions need an authenticated session on top of the word, that closes off most of the casual-misuse angle. Voiceprint and replay protection sound like the last real gap between this and something I'd trust with a production repo. Good luck closing it out.
Dial
this solves a real problem, I live in Claude Code all day too and the "still tied to a laptop to babysit it" gap is exactly right. question that's more about reliability than the approval-safety angle everyone's asking about: what happens if you lose connectivity mid-flow, say you're walking and your phone drops signal right as an agent is waiting on a voice approval. does it just pause and wait indefinitely for you to reconnect, timeout and roll back the pending action, or keep going with whatever the last confirmed state was?
Openbase
@galdayan If your phone drops right as an agent is waiting on approval, the agent stays blocked at that approval checkpoint. The pending request is stored in the local approval queue on your Mac, and when the phone reconnects it shows back up in the Approvals tab.
Already completed work isn’t rolled back, but the gated action also does not run unless an approval actually made it through. If the approval was recorded before the signal dropped, the agent continues from that confirmed state. If not, it waits.
Creatium
Openbase
@ridhwikvinod Oh yeah I do the same haha. In general, Openbase doesn't give an interpretation by default; it's pretty good at following along despite corrections or half sentences. But you can always ask at the end of your prompt for Openbase to confirm what it understood from your prompt, and we're also adding a system instructions setting as well so that can be something you can put in there.
@ridhwikvinod @lucaszhao Exactly. And once you save that instruction as a Skill, the confirmation can become part of the workflow instead of something you add to every prompt. For example: “Always summarize the task back to me before starting any long-running work.”
Honestly the voice steering thing is kind of wild, I was telling an agent to fix a bug while walking the dog and it just committed the patch. Felt like cheating honestly.
@dilaraekme8ruh Haha, that's exactly the feeling we were hoping for 😄
I'm non-technical and ship almost everything through coding agents, so the part that would decide it for me isn't approval, it's verification. My recurring failure isn't an agent doing something destructive, it's an agent telling me a change shipped when it actually landed on a branch that had already been merged. Does Openbase report back from the actual repo state after the fact, or only from what the agent says it did?
Openbase
@podcast_ai Good question! We've been working on a separate tab purely for git trees etc. to visualize and clearly see diffs and what branch you're on. Currently it is still agent reported, but we'll definitely implement this soon!
This is the hard part and you're putting it front and center: which calls the agent makes alone vs. which ones it asks about out loud.
We ship agents in a very different domain — they message customers and take payments for small businesses — and the rule we landed on after getting it wrong is: plan freely, execute visibly, but always ask before anything that touches money, anything that goes out to many people at once, and anything that changes configuration.
The trap we hit: we tried letting the agent tune its own thresholds, with a guard scoring the outcomes. It optimized straight into the failure mode — a safe, wrong answer scores beautifully, because it generates no escalation and burns fewer tokens. We killed the autonomy and kept the human in the loop.
Curious how you think about that when the approval is by voice, where "yes" is a very cheap thing to say.