Skip to main content

​
🚀 Autoplay in one breath

Autoplay streams what users are doing in your product into your agents as clean, LLM-ready context — so copilots can help before someone asks, and stay quiet when they shouldn’t interrupt. We run the connector pipeline (extract, normalise, optional summarise); you wire callbacks and RAG.

​
⚡ What Autoplay handles for you

  • Real-time events at scale — browser activity becomes normalised, typed payloads your model can read
  • Context tooling — buffers, summarisers, and models so volume does not blow your context window
  • Golden paths — record ideal journeys with the Autoplay Chrome extension or dashboard
  • Workflow completion — per-user mastery, in-progress, and gaps across sessions so suggestions stay relevant

​
🔗 How it fits in a RAG system

autoplay-sdk consumes your connector over SSE (or push webhook) and hands you typed Python objects; you embed, upsert, and retrieve like any other RAG source. 
Real-time user events
      ↓
Autoplay connector pipeline
  (extract → normalise → summarise)
      ↓
SSE stream  ←──── autoplay-sdk ────→  your callback
                                           ↓
                                     embed + upsert
                                           ↓
                                      vector store
                                           ↓
                                    RAG retrieval / chatbot

​
📦 Two event types

actions

A batch of UI interactions extracted from a user session — page views, clicks, inputs. Best for granular, per-session embeddings.

summary

An LLM-generated prose summary of a session, replacing the raw action list. Best for compact context windows in RAG pipelines.

​
🏷️ Typed models

All callbacks receive typed dataclass instances — no raw dict parsing in your code.
ClassEvent type.to_text() output
ActionsPayloadactionsNumbered list of actions, embedding-ready
SummaryPayloadsummaryProse summary string directly
SlimAction—"{description} — {canonical_url}"

​
🔌 Two clients

ClientUse when
ConnectorClientYour pipeline is synchronous (no async/await)
AsyncConnectorClientYour pipeline uses asyncio — LangChain, LlamaIndex, FastAPI

​
📡 Two delivery modes

ModeHow it worksSDK class
SSEYour service connects to the connector streamAsyncConnectorClient / ConnectorClient
Push webhookConnector POSTs to your endpointWebhookReceiver
Both modes deliver the same typed ActionsPayload / SummaryPayload objects.

​
🗄️ Two buffer options

BufferStorageUse when
EventBufferIn-memoryDevelopment, single process
RedisEventBufferRedis ZSET (sliding window)Production, multiple pods

​
Install

pip install autoplay-sdk

from autoplay_sdk import (
    # Clients
    ConnectorClient,
    AsyncConnectorClient,
    # Typed models
    ActionsPayload,
    SummaryPayload,
    SlimAction,
    # Buffers
    EventBuffer,
    RedisEventBuffer,
    BufferBackend,
    # Push webhook receiver
    WebhookReceiver,
)