What is Rivet?
Rivet is an open-source orchestrator for agentic workloads built around long-lived processes called Actors. Rivet combines application state, durable storage, real-time networking, and scheduling in an Actor that developers can assign to an agent, user, session, or tenant. Rivet also supports general stateful applications rather than requiring an AI agent use case.
Rivet separates the agent loop from the sandbox used to execute generated code. Agent memory, state, and context live in the Actor, while a sandbox can be reached through tools when code execution is necessary. This architecture addresses agents that spend much of their time waiting for inference rather than actively using compute.
What Rivet can do
Rivet provides application primitives for durable execution and live client interaction. Each Actor combines in-memory variables with a SQLite database, while networking and scheduling support ongoing sessions and background work. Rivet also ships libraries for agents, workflows, and generated applications, so developers can build directly with Actors or use a library for a particular workload.
- Rivet stores agent transcripts and artifacts in durable SQLite databases.
- Rivet broadcasts real-time events to connected WebSocket clients.
- Rivet hibernates idle Actors and wakes them on the next request.
- Rivet moves Actors to another node after a node failure while preserving state.
- Rivet schedules Actor wake-ups through cron and Actor-defined times.
- Rivet supports Actor-to-Actor calls across a cluster.
- Rivet exposes traces, metrics, and logs in OpenTelemetry format.
How Rivet works
Rivet applications define Actors and expose actions that clients can call. The repository's TypeScript example creates an agent with SQLite message history, in-memory presence tracking, and streamed responses broadcast over WebSockets. A frontend or backend client gets or creates the agent, connects to the Actor, subscribes to events, and invokes actions.
Rivet runs on native Node.js and Bun without a custom runtime and provides typed clients and React hooks. Developers can work locally before deploying the control plane through Docker, Kubernetes, or a virtual machine. Production storage can use Amazon S3, Google Cloud Storage, Azure Blob Storage, or an S3-compatible object store.
Who Rivet is for
Rivet serves developers building applications that need persistent identity, shared state, and live connections. Agent developers can preserve memory across restarts, while application developers can assign Actors to documents or users. Rivet also supports agents that only call tools, MCP, and APIs, without requiring a sandbox for those interactions.
- Rivet supports multiplayer AI agents with memory shared live across clients.
- Rivet supports multi-step workflows that replay instead of starting over.
- Rivet runs applications and Actors generated by an agent.
- Rivet supports collaborative editing with an Actor per document.
Rivet pricing and deployment options
Rivet is open source under the permissive Apache 2.0 license and can be run locally or self-hosted. The repository does not provide prices for managed services. Rivet Cloud offers fully managed, dedicated clusters on AWS or GCP, with either hosted nodes or nodes supplied by the customer.
Rivet also offers BYOC deployment with the control plane and storage inside the customer's VPC on AWS, GCP, or Azure. The repository states that BYOC supports air-gapped environments and keeps data inside the customer's cloud. Every deployment option uses the same control plane, from local development through production.
Rivet strengths and limitations
Rivet combines durable storage, real-time communication, scheduling, and fault tolerance within the Actor model. Self-hosted deployments need an object storage bucket rather than a separate database or message bus. However, self-hosting still requires infrastructure configuration, and Actors do not replace a sandbox when an agent needs to execute generated code.
- Rivet self-hosting requires operating the control plane and configuring its storage backend.
- Rivet agents that execute generated code need an appropriate sandbox.
- Rivet Dynamic Apps is currently described as a preview.
Rivet FAQ
What is Rivet?
Rivet is an open-source orchestrator built around long-lived Actors with in-memory state, SQLite storage, WebSockets, and scheduling. Rivet supports durable AI agents, replayable workflows, collaborative editing, and other stateful applications.
Is Rivet open source?
Rivet is open source under the permissive Apache 2.0 license and supports local use and self-hosting. Rivet also offers managed Rivet Cloud and BYOC deployments, but the repository does not state prices for those services.
Can Rivet be self-hosted?
Rivet can be self-hosted on Docker, Kubernetes, virtual machines, or bare metal. Rivet requires object storage rather than a separate database or message bus, and supports Amazon S3, Google Cloud Storage, Azure Blob Storage, and S3-compatible stores.
Does Rivet require a sandbox for AI agents?
Rivet does not require a sandbox for agents that only use tools, MCP, and APIs. Rivet agents need a sandbox when executing code they write, using agentOS or an existing sandbox provider exposed through tools.
What integrations and runtimes does Rivet support?
Rivet runs on native Node.js and Bun without a custom runtime, and the repository also lists Rust among supported runtimes. Rivet provides typed clients and React hooks and lists integrations with Effect, Vitest, OpenTelemetry, Pino, AI SDK, OpenAPI, and AsyncAPI.
What happens when a Rivet Actor becomes idle or a node fails?
Rivet hibernates idle Actors and wakes them on the next request without losing state. Rivet also moves Actors to another node with their state intact when a node goes down.
From the community
No reviews yet
Be the first to review this tool. What you write helps the next person choose.




