Durable State

Durable State adds structured state changes on top of Durable Streams. Instead of treating a stream as raw bytes, you work with typed insert, update, and delete events.

Deploy now ↗

Protocol overview

Durable State uses JSON streams (Content-Type: application/json) and define two message types:

• Change messages for insert, update, and delete
• Control messages for snapshot boundaries and resets

Clients append these events to a stream and materialize state by applying them in order.

Use Durable Streams directly for raw token or byte streaming. Use Durable State when you want database-style sync semantics on top.

See the full State protocol specification for the formal wire format and requirements.

Installation

npm install @durable-streams/state @tanstack/db

@tanstack/db is only needed for StreamDB. If you only need MaterializedState, you can skip it.

Change events

The State Protocol defines a standard format for state change events. Each event targets a typed entity identified by type and key, and carries an operation in its headers:

{
  "type": "user",
  "key": "user:123",
  "value": { "name": "Alice", "email": "alice@example.com" },
  "headers": {
    "operation": "insert",
    "txid": "abc-123",
    "timestamp": "2025-12-23T10:30:00Z"
  }
}

Fields:

Field	Required	Description
type	Yes	Entity type discriminator -- routes events to the correct collection
key	Yes	Unique identifier for the entity within its type
value	For insert/update	The entity data
old_value	No	Previous value, useful for conflict detection
headers.operation	Yes	One of "insert", "update", or "delete"
headers.txid	No	Transaction identifier for confirmation
headers.timestamp	No	RFC 3339 timestamp

Multiple entity types coexist in the same stream. A chat room stream might carry user, message, reaction, and typing events, all interleaved and processed in order.

Control events

The protocol also defines control events for stream management, separate from data changes. These have a control field in their headers instead of an operation:

Control	Purpose
snapshot-start	Marks the beginning of a snapshot -- a complete dump of current state
snapshot-end	Marks the end of a snapshot boundary
reset	Signals clients to clear their materialized state and restart

{"headers": {"control": "snapshot-start", "offset": "123456_000"}}

{"type": "user", "key": "1", "value": {"name": "Alice"}, "headers": {"operation": "insert"}}
{"type": "user", "key": "2", "value": {"name": "Bob"}, "headers": {"operation": "insert"}}

{"headers": {"control": "snapshot-end", "offset": "123456_789"}}

You'll encounter these when a server sends a full state snapshot rather than incremental changes -- for example, on initial connection or after a schema migration.

MaterializedState

MaterializedState is a simple in-memory key-value store that applies change events. It's the minimal way to consume state protocol events -- no schemas, no reactive queries, just a map from (type, key) to the latest value.

import { MaterializedState } from "@durable-streams/state"

const state = new MaterializedState()

state.apply({
  type: "user",
  key: "1",
  value: { name: "Alice" },
  headers: { operation: "insert" },
})

state.apply({
  type: "user",
  key: "1",
  value: { name: "Alice Smith" },
  headers: { operation: "update" },
})

const user = state.get("user", "1") // { name: "Alice Smith" }
const allUsers = state.getType("user") // Map of all users

MaterializedState is a good fit when you need straightforward state tracking without reactive queries or schema validation.

StreamDB

For applications that need reactive queries, filtering, joins, and optimistic updates, use StreamDB.

StreamDB is the createStreamDB / StreamDB layer in @durable-streams/state/db, built on top of State Streams and TanStack DB.

Durable Sessions

A Durable Session multiplexes AI token streams with structured state into a persistent, shared session. Multiple users and agents can subscribe to and join the session at any time, making it a natural fit for collaborative AI applications.

The pattern layers protocols on top of each other:

1. Durable Streams -- reliable, resumable byte delivery
2. State Protocol -- structured CRUD operations over streams
3. Application protocols -- AI SDK transports, presence, CRDTs

This enables scenarios where an AI agent streams tokens into a session while structured state (tool results, user presence, shared documents) flows through the same infrastructure.

For a detailed walkthrough, see the Durable Sessions for Collaborative AI blog post.

Learn more

• State protocol specification -- full protocol spec
• StreamDB -- reactive collections, queries, and optimistic actions
• Yjs -- sync Yjs CRDTs for collaborative editing
• Package README -- complete API reference
• Examples -- background jobs dashboard and Wikipedia live events demo
• TanStack DB -- reactive collections and query engine
• Standard Schema -- schema validation

---

See also: Core concepts | JSON mode | StreamDB | Yjs