Install for free. Pay for the work the agent does.
Three ways to run a specialist inside your Lovable app. No seats, no per-user licence, and nothing to commit to before you have seen it work.
How the meter works, in one sentence: we count the text the agent reads and the text it writes, added together, in the units the industry calls tokens — a short customer request and its answer is a few hundred of them.
Browse the hub and its reference checksTry it
Trial tasks on us, so you can see what it does before anything is billed.
- 100k tokens a month
- Shared capacity, so answers can queue at busy times
- Help through the community forum
Pay for the work done
You pay for what the agent reads and writes, at one rate, billed monthly. Nothing to commit to.
- One rate for the text going in and the text coming out
- No monthly minimum
- Help by email
billed monthly by usage
Hire a dedicated specialist
An agent trained on your domain, with a reference check run on your own work.
- Trained on your tools, your request shapes and your rules
- Its reference check is scored on a set of your own tasks that we lock away before training starts — the agent never sees them
- Private deployment available
- A named contact at Blade while it is in service
How hiring a dedicated specialist works
Your tool inventory and a slice of representative traffic — enough for us to see the shapes of the work your team actually handles.
Before any training begins we set aside a sample of your tasks. The agent never sees them, and they are what its reference check is scored on. This order is not negotiable, because a test chosen after training is not a test.
Our factory builds a training corpus for your domain and tunes an agent on it. This takes days, not quarters.
You get its scores on your own locked-away tasks — including whatever it does badly — and a private endpoint to run it against.
About that traffic: it stays on machines we own, it is used to build and score your agent, and nothing in it is used to train anything else.
You pay for what is actually used. We publish no comparison between our agent and any other model: the one comparison we did publish was withdrawn on 2026-07-30, because the two arms were not run as a controlled comparison, and we have no replacement we would defend. What we do publish is our own agent's reference check, including the duty it missed.
Read the reference check