Rivet, an open-source project on GitHub under rivet-dev/rivet, describes itself as an orchestrator for agentic workloads that schedules Actors instead of containers¹.

The repository, licensed Apache 2.0, says each Actor is a long-lived process with durable state, a SQLite database, a queue, realtime connections, and scheduling, addressed by key¹.

According to the README, the control plane is a single Rust binary plus an object-storage bucket; it says there is no database to run and no message bus, and that control-plane instances find each other through the bucket¹.

The maintainers say new, untrusted code generated at runtime can run in its own Actor, with its own database and filesystem and no access to others'¹.

The repo lists typed SDKs for backends and React hooks for frontends, and says there is no custom runtime so existing packages, tooling, and tests can be used unchanged on Node.js and Bun¹.

The README lede cites 6.9 ms cold starts and 72 KB per Actor. Benchmark notes dated October 25, 2026 on Azure eastus2 report 6.92 ms p50 cold start (8.11 ms p99, 1,500 samples) and 4.37 ms p50 wake from hibernation of at least 10 seconds (6.80 ms p99, 1,200 samples), with 72.4 KB RSS increase per running Actor measured with the Rust rivetkit SDK¹.

Those numbers are the project's own benchmarks; independent verification is not in this source¹.

Deployment options listed include local and self-host, a hosted Rivet Cloud, bring-your-own-nodes with Rivet providing control plane and routing, and BYOC inside a customer's VPC¹.

The monorepo includes RivetKit TypeScript, Rust, Python and Swift client trees, engine code, a dashboard, documentation, and examples¹.

What is not established in the GitHub page is a dated product launch announcement, Y Combinator status, pricing, or third-party adoption¹.

Sources

1. github.comrivet-dev/rivet12:17 PM

Read the original at github.com ↗