Rivet Open-Sources Agent Orchestrator With 5ms Cold Starts

Rivet, a YC F25-backed startup, officially launched on November 6, 2026, with an open-source actor orchestrator designed for agent workloads. The platform claims 4.99ms p99 cold starts and a 72KB memory baseline per actor, positioning itself as a lightweight alternative to sandbox-based approaches and Cloudflare Durable Objects. Released under Apache 2.0, the system pairs long-lived in-memory state with SQLite databases backed by Amazon S3.

Architecture and Performance Claims

The Rivet orchestrator builds on an actor model where each actor functions as a long-lived process with in-memory state, a WebSocket server for multiplayer scenarios, and durable SQLite storage. Founders Nathan Flurry and Nicholas, who previously built real-time infrastructure for multiplayer games Diep and Krunker, migrated the platform from FoundationDB to an S3-backed design that acknowledges writes once SSDs in two of three availability zones confirm, then batches to S3 asynchronously.

According to the GitHub README and Launch HN post, Rivet delivers 1.07ms durable writes and 1.36ms reads at p99. The team reports that actor creation runs 39x faster, SQLite writes 28x faster, and backend requests 13x faster than Cloudflare Durable Objects. A minimal sandbox approach, by comparison, reportedly requires 1.7 seconds and 512MB of memory per instance. The storage layer runs 10.4x cheaper than replicated SSD storage, the founders claim.

Direct Comparison to Cloudflare Durable Objects

Rivet positions itself as a frontal alternative to Cloudflare Durable Objects, removing the 128MB memory cap and 10GB SQLite limit while eliminating random actor evictions. The platform runs on vanilla Node.js or Bun instead of a custom runtime, which the team says reduces friction for developers migrating existing agent codebases.

All performance numbers are self-published. Neither the HN post nor the README specifies the test workload, hardware configuration, or dataset used for the benchmarks. The project launched with 6.7k GitHub stars and lists one named migration customer, agent startup Amp, which reportedly moved on-premises infrastructure over a weekend.

Storage Layer and SDK Support

The S3-backed storage architecture is central to Rivet's cost and latency pitch. By batching writes asynchronously after confirming across availability zones, the platform reduces storage costs while maintaining durability guarantees. Rivet ships SDKs for TypeScript, Rust, Python, and Swift, covering the primary languages used in agent development.

Each actor supports hibernation when idle, reducing memory overhead for inactive processes. The platform is designed to scale to billions of actors, a claim tied to the low per-actor memory footprint and the decoupled storage model. The GitHub repository provides deployment guides for self-hosted and cloud environments.

Release Date and Availability

Rivet was officially launched on November 6, 2026, with full documentation and open-source code available on GitHub under Apache 2.0. The platform is immediately available for self-hosting and cloud deployment. The founders have not announced a managed service tier or enterprise support plan, though the open-source license permits commercial use without restriction.

What This Means

Rivet enters a crowded agent infrastructure market where cold start latency and memory efficiency are recurring pain points. The sub-5ms cold start claim, if validated in production workloads, represents a meaningful step down from typical sandbox initialization times. However, the lack of independent benchmarks and the single named migration customer suggest early-stage adoption. The S3-backed storage model trades off immediate consistency for cost efficiency, a tradeoff that will appeal to price-sensitive agent deployments but may raise questions for workloads requiring strict transactional guarantees. The Apache 2.0 license and multi-language SDK support lower barriers to experimentation, positioning Rivet as a pragmatic option for teams building stateful agent systems at scale.