Chat / Sessions

Durable, real-time messaging for AI agents and human conversations.

What it is

Chat gives every conversation a persistent thread backed by the Durable Objects ring. Messages survive restarts. Token streams from LLMs accumulate in memory and flush to subscribers every 40ms. When a run ends, the final message is committed to the WAL and Postgres.

Mental model

A thread is the durable unit — one WebSocket channel per thread:

chat:v1:app:{appKey}:thread:{threadId}

A run is one LLM generation: it starts, streams tokens, and ends. Token chunks bypass full state serialisation via set_volatile — only a tiny log entry is written per flush, not the entire thread state.

Run lifecycle

When to use

• AI chat interfaces — customer support, copilots, assistants
• Multi-agent pipelines that need a durable audit log
• Any product with persistent, searchable conversation history

If you only need fast pub/sub without persistence, use Channels instead.

New here? Start with the Quickstart — you'll have a working chat interface in 15 minutes.

Know what you're looking for? Jump to the Reference.