What is Rivet (rivet.dev)?

Updated

Rivet is an open-source, self-hostable orchestrator for agentic workloads. It runs agents, workflows, and sandboxes as durable Actors, so each agent gets a persistent process with state that survives restarts. It's aimed at developers building AI agent applications who need durable execution, high density, and fast cold starts — and who want to run it in their own infrastructure.

What Rivet actually does

The core abstraction is the Actor: a durable process that an agent "lives in." Instead of treating each agent run as a stateless request, Rivet keeps it alive as a long-running unit that can be resumed. On top of that, Rivet offers three workload types:

Workload What it provides
Actors A durable process for each agent to live in
Workflows Multi-step operations that replay instead of starting over
Sandboxes A filesystem, shell, and network for code you did not write

There are also higher-level patterns built on these: Dynamic Apps (a backend per user, deployed the moment it's generated), agentOS for coding agents, and Enterprise deployments where the VPC keeps data inside your own cloud.

The key mechanism: Actors, not containers

The distinguishing design choice is where an Actor runs. An Actor lives inside a worker process you already run, not in its own container or VM. Rivet's site claims this density advantage directly:

  • A server that held hundreds of pods can hold hundreds of thousands of Actors.
  • 72.4 KB per running Actor (RSS increase per Actor).
  • 12 ms cold starts.
  • "Run 2,829× more workloads per server."

That's the practical difference from ordinary container or VM orchestration: you're not paying the per-container overhead for every agent. For workloads that spin up many short-lived or per-user agents, that overhead is usually the bottleneck.

Who it's for

Rivet fits developers who are:

  • Building AI agent applications that need persistent state and durable execution.
  • Running many agents per server and hitting density or cold-start limits with containers.
  • Needing sandboxed execution for code they didn't write (a common agent requirement).
  • Wanting open-source and self-hostable infrastructure rather than a managed-only platform.
  • Deploying per-user backends or company-specific agents inside their own VPC.

If you're running a single long-lived service, or you don't need durable state across restarts, the Actor model adds concepts you may not need.

How to start

Rivet's site points to a coding-agent skill install as the entry point:

npx skills add rivet-dev/skills

The site describes this as something you "give to your coding agent," which suggests the intended workflow is agent-assisted setup rather than manual configuration. Beyond that, the site links to separate exploration paths for agentOS, Dynamic Apps, Enterprise, Actors, and Realtime apps.

What to verify before committing

The site's headline claims (12 ms cold starts, 72.4 KB per Actor, 2,829× density) are vendor-stated figures, so treat them as targets to validate against your own workload rather than guarantees. Pricing is not specified in the available material — there is a pricing page, but no plan details or payment information is given, so don't assume it's free or that no login is required. If cost or hosting model matters to your decision, check the pricing page directly before building on it.