Launched this week

Teaflask
Self-improving AI copilot for your app. No backend changes.
259 followers
Self-improving AI copilot for your app. No backend changes.
259 followers
Turn your existing SaaS into an AI-native product without rebuilding your backend. Teaflask adds a managed AI copilot that lets users get things done by simply asking. Integrate through your frontend, use your own UI components, and go live in a day. Teaflask keeps its product knowledge grounded in your code, learns from real usage, and lets you test changes before release.










Free Options
Launch Team

Transformer.gg by Jon Oringer(founder of Shutterstock)The agentic influencer marketplace for brands, influencers, and bots.
Promoted
api calls going through the user's own browser session is the smart bit, no second permission system to keep in sync with teh real one
how does Ask work in practice, does the end user confirm each action or does an admin decide once for everyone??? and when you replay old conversations on a cheaper model, what counts as a pass???
Thanks @thamibenjelloun ! Yeah, that was precisely the idea. Your app already knows what each user is allowed to do, so we didn’t want to build a second copy of that.
On Ask, its actually both. You, as the admin, choose what should require approval, and that rule applies to everyone. If an action is marked to Ask, the user gets a card into the chat confirming exactly what is going to be called, and the user must approve the call before anything actually occurs.
And as for what is a pass, we convert each of our saved conversations into a test case, simulating the user end of the conversation, then replaying the original responses from the API (so we don't make any calls out against live customer data)
.
There are basically three things we check:
- We verify in the code that it made the correct calls, used the correct parameters, and that those calls were successful. We verify it didn't use actions it shouldn't have used.
- Did it do the right thing? The LLM judge verifies things that are tricky to write down in code. For instance, whether it inquired about the specific project the user was referring to before deleting things. We create these checks from the initial conversation, but you can modify them.
- Did it actually finish?
We run each case several times, and all runs must pass.