The tools are genuinely capable when they work. The work assistance, the document integrations, the ability to hold a complex idea across a long project — that part is real and occasionally impressive - if and when it works. Review collected by and hosted on G2.com.
The pitch is compelling: an AI agent that remembers your work, builds on your context, and gets things done without you having to repeat yourself. The reality is that every session begins with the agent rediscovering what it already knew, searching in the wrong places for material it was explicitly told how to find, and burning your paid credits on the loops.
I'm working on a complex things. I've uploaded source documents, built a database, and spent hours establishing context. None of that prevents the agent from ignoring the database entirely and running five redundant searches before landing on the file that was sitting there the whole time. When I pointed this out, I got an apology and a promise to do better — then the same thing happened the next session.
But the amnesia is only half the problem. The other half is that the agent has no reliable working knowledge of its own platform. It will confidently begin building a solution — a database structure, a backend function, an integration — and only discover mid-build that it's own source platforms and API don't support what it's doing, the feature has a limit it didn't know about, or the entire approach was never going to work. It also routinely gives instructions to go to features in the Base44 platform that don't exist.
By then you've spent thirty messages and a meaningful chunk of your monthly credits watching it construct a dead end. It doesn't just waste the credits. It fills the agent's working memory with the wreckage of the failed attempt, crowding out the actual project context you spent hours establishing.
The tools are genuinely capable when they work. The writing assistance, the document integrations, the ability to hold a complex argument across a long project — that part is real and occasionally impressive. But you pay per message, and a significant percentage of your messages are the agent learning things it should already know, either about your project or about itself.
Base44's response to platform feedback appears to be SOL, especially about credits. This would be a 4-5 star review if they didn't make their users finance all this, but because compute is being used, your paying for all of this, like it or not.
You're also paying for any bugs. Example, the agent has been asking me for days to approve things using an approve button in the UI that has been broken for days, and doesn't work.
When the agent can't act without your approval and the approval button doesn't work, you are simply stuck.
What you are paying for is a super-powerful tool that has something along the lines of the digital equivalent of a cognitive disorder.
It is capable of extraordinary work in short bursts, and then it loses the thread, rebuilds something that was already built, discovers a wall it should have known was there, and charges you for all of it.
If you have simple, self-contained tasks, this might work for you. If you're doing deep, ongoing creative work that depends on continuity and coherent execution — the thing they're specifically selling — manage your expectations carefully.
btw, my agent wrote this about itself, based on our shared experience. Go figure. Review collected by and hosted on G2.com.
