Use case · Evals and experiments
Replay any run from the same state.
Agent evals are only fair when every run starts from the same world. Snapshot it once, fork it per run, and throw the forks away.
How it fits
Built from the primitives you already get.
Golden snapshots
Capture a database, filesystem or memory as a named restore point and start every run from it.
A fork per run
primitive_fork copies a primitive for each model, prompt or tool variant, so runs never see each other's changes.
Inspect the outcome
Query each fork's database or search its files to grade what the agent actually did, not just what it said.
Run it in CI
The open-source engine runs locally with primitives dev, so evals run in CI without a cloud account.
Build
Local-first development.
Build and test on your machine with the open-source engine. No cloud account needed: every primitive runs locally, with the same API as the cloud.
Install the CLI, then run primitives dev. Quickstart
Open source
Built in the open. Contribute freely.
The SDKs, the MCP server and the single-node engine are open source and actively maintained. Submit PRs, report issues, improve docs, or just star the repo to show support.
useprimitives/primitivesFork, star, or open a pull request