Build the agent, skip the year underneath it.
Your product is the agent. The quarter goes to the machine per session, the credentials per user, the memory that must not cross users, and the meter that says what each user costs you. Levain gives every user a workspace with all four in it.
Your users never need a Levain account.
A machine per run.
A fresh microVM, gone when the run ends. Nothing a user's session writes outlives it.
What one user tells the agent stays with that user.
Memory and credentials live in the user's workspace and are readable only from inside it.
You find out which user is expensive.
Credits pool at your organization; every workspace reports its own share of the month, and you can cap it.
Onboard a new user with one call.
The same four steps whether it is your first user or your thousandth.
Write the agent once.
Python with the SDK, or any language through the API. Publish it as a version in your registry.
Open a workspace per user.
Sandbox, knowledge, memory and limits come with it. Nothing to assemble.
Hand over the user's connections.
Their accounts, their files, their data. The credentials stay in their workspace, and the model never sees them.
Publish.
Every workspace that adopted the agent follows the new version. Read what each user used; suspend the ones that leave.
The calls themselves, with every field. Read the docs
Instead of: a sandbox API, a queue, a memory API and a spreadsheet of who owes what.
Today your agent runs on our loop, not yours.
Bringing your own loop is on the roadmap; it is not in the product. If you have a loop you will not give up, talk to an engineer, because we are looking for the first team to build it with.
Frequently asked questions.
I already have an agent loop. Can I bring it?
Not today; the agent runs on Levain's own agent loop inside the workspace. Running the runtime and isolation under a loop you built is on the roadmap, and we would like to build it with you.
What isolates one user's run from another's?
Each user has a workspace with its own sandbox, credentials, knowledge base and memory, and an agent can only read the workspace it runs in. Every run starts in a fresh microVM that is gone when the run ends.
What does a user cost me?
Credits are pooled at your organization and every workspace reports its month-to-date usage. You see and cap what each user spends.
Which models can I use?
Fourteen providers through one setting, Anthropic, OpenAI, Google, Meta, Mistral and nine more, per task. Or your own keys.
Open your first user's workspace today.
A free organization comes with a workspace to build in. Add your first user when the agent is ready.