Durable Streams is an HTTP-based protocol for ordered, replayable data streams that client applications can read from any offset and tail live. Each stream is a URL, so a browser tab, mobile app, native client, IoT device, or edge worker can disconnect, refresh, or switch devices and resume exactly where it left off, and several viewers can watch the same stream at once.
It targets the gap between fragile WebSocket or SSE connections and backend-only durable logs such as database write-ahead logs and Kafka topics. In AI apps, that means LLM token streams, tool outputs, and progress events survive reconnects without re-running the model call. Other uses include database sync to clients, collaborative editing with CRDTs, event sourcing, and workflow state history.
The protocol grew out of Electric's real-time Postgres sync. Reads are offset-based, live tailing uses long-poll or SSE, payloads are content-type agnostic byte streams (with a JSON message mode), and the HTTP-native design lets a CDN fan one origin out to many concurrent viewers. The project is in beta.
Client libraries exist for TypeScript, Python, Go, Elixir, .NET, Swift, PHP, Java, Rust, and Ruby. Servers include a Node.js reference server for development and a Caddy plugin for production, alongside a CLI and server and client conformance test suites.
Features
- Resumable reads: offset-based catch-up from any point, plus live tailing over long-poll or SSE
- Addressable streams: each stream is a URL that several clients can watch together in real time
- JSON mode: stream structured data as JSON messages
- Durable Proxy: a durable proxy for AI token streams so a disconnect does not lose or re-run a response
- Durable State: a state protocol for insert, update, and delete over streams
- StreamDB and StreamFS: a type-safe reactive database and a synced filesystem built on a stream
- AI framework integrations: a Durable Transport adapter for AI SDK apps and Durable Session support for TanStack AI
- Collaborative editing: a Yjs provider for collaborative documents
- Multi-language clients: TypeScript, Python, Go, Elixir, .NET, Swift, PHP, Java, Rust, and Ruby
- Servers and conformance: a Node.js reference server, a Caddy server plugin, and protocol compliance test suites
