Pay for agents that work.

Any call that touches a primitive's data or lifecycle through REST, the SDKs or MCP: a query, exec or batch, a file read or write, a queue send or receive, a memory recall, a secret set or fetch, a key get or set, an object put or download (including through a signed URL), a snapshot, fork, restore, pause, resume or export. Every MCP tool call on a primitive counts, including each call in a JSON-RPC batch. Creating primitives, listing groups and primitives, managing API keys, MCP initialize and tools/list, and dashboard page loads are free. Queue messages are just operations, not a separate limit.

Nothing is billed. After 1M operations in a month, further operations are refused with 429 plan_limit until the counter resets on the 1st of the month (UTC). At 5 GB of storage, writes are refused with 429 plan_limit while reads and deletes keep working. Starter holds 1 group and up to 100 primitives; creating a 101st is refused the same way. Upgrade to Pro any time to lift every cap right away.

No. Pro includes 25M operations and 50 GB of storage each month. Usage past that is billed as overage at $0.50 per extra 1M operations and $0.25 per GB-month, so your agents keep running.

Ten times the included usage, 250M operations and 250 GB each month, with lower overage at $0.40 per extra 1M operations and $0.20 per GB-month. The audit log goes back 90 days instead of 30, your team signs in with SAML SSO, and support moves to priority email plus a shared Slack channel.

Yes, on Enterprise. With bring your own cloud (BYOC), Primitives is deployed into your own cloud account, so your agents' data never leaves your infrastructure, and we take care of upgrades and operations. Enterprise can also run on dedicated infrastructure in the region you choose. Talk to us to scope it. To run everything yourself without a contract, the open-source engine is free to self-host.

You could wire those up yourself, but you'd be managing three services, three sets of credentials and your own isolation for every agent. Primitives puts all seven behind one API and one key. Each agent gets as many databases, filesystems, queues, memories, vaults, key-value stores and object stores as it needs, and you can snapshot, fork or pause any of them on its own. You also get a single audit trail and see what each agent costs.

No. Every primitive is exposed as MCP tools, so any MCP-capable agent can use them directly. The SDKs, the REST API and MCP all run through the same API.

No. Agents reference secrets by name. The engine injects the value into an outbound HTTPS request that matches the secret's grant, redacts it from the response, and never returns it in a tool result.

Storage only. An idle group is just its primitives' files and makes no operations; the next call opens them.

Yes. The engine is open source: primitives dev runs the REST API and MCP on localhost, with each group stored as a folder of SQLite files. Self-hosted, every feature is included: snapshots, forks and unlimited groups. The cloud runs the same engine and charges for running it for you, so code moves between them by changing the API key and URL.

Each API key is scoped to one group or to the whole workspace, and can optionally be limited to some tools, such as read-only database queries. Give each agent its own key scoped to its group, or let MCP clients sign in with OAuth, so your workspace key never leaves your server.

Any sandbox that can make HTTPS calls: E2B, Daytona, Modal, Firecracker or your own containers. Code and agents inside call the API or MCP with an API key scoped to one group. Filesystem mounts and sandbox adapters are coming soon.