Integrations

Every provider threads ships an adapter for, what to import it from, and where it is documented.

Each integration is a factory function that returns an adapter: a model, a sandbox, a channel, a memory provider, a search backend, a store or an exporter. You pass it to an agent or a host; the framework never talks to a provider any other way.

Adapters live in their own package (TypeScript) or behind their own extra (Python), so the core depends on nothing but zod and pydantic. Installation lists the packages and extras. Which of these each language has today is in Python and TypeScript.

Models

	TypeScript	Python	Guide
Anthropic	`anthropic()` from `@threads/anthropic`	`anthropic()` from `threads.anthropic`	Models
OpenAI	`openai()` from `@threads/openai`	`openai()` from `threads.openai`	Models
Everything else	`aiSdk()` from `@threads/ai-sdk`, for any AI SDK provider	`litellm()` from `threads.litellm`, over LiteLLM's OpenAI-compatible route	Models
Scripted model, for tests	`scriptedModel()` from `@threads/core`	`scripted_model()` from `threads`	Testing

Sandboxes

	TypeScript	Python	Guide
E2B	`e2b()` from `@threads/e2b`	`e2b()` from `threads.e2b`	E2B
Daytona	`daytona()` from `@threads/daytona`	`daytona()` from `threads.daytona`	Daytona
Modal	—	`modal()` from `threads.modal`	Modal
Fake sandbox, for tests	`fakeSandbox()` from `@threads/core`	`fake_sandbox()` from `threads`	Testing

Modal is Python only for now: modal() from @threads/modal throws ConfigError transport_fence_unsupported, because Modal's JavaScript SDK sends commands over a connection threads can't safely wrap.

Credentials stay on the host. A sandbox never receives a key, and has no internet unless you give it one. See Sandboxes.

Channels

	TypeScript	Python	Guide
Slack	`slack()` from `@threads/slack`	`slack()` from `threads.slack`	Slack
WhatsApp	`whatsapp()` from `@threads/whatsapp`	`whatsapp()` from `threads.whatsapp`	WhatsApp
GitHub	`github()` from `@threads/github`	`github()` from `threads.github`	GitHub

Channels run under the host, which also serves the HTTP API and cron schedules.

Tools from other servers

	TypeScript	Python	Guide
Any MCP server	`mcp()` from `@threads/mcp`	`mcp()` from `threads.mcp`	MCP servers

Web search

The web_search tool needs a backend. All three come from the core package.

	TypeScript	Python	Guide
Exa	`exa()` from `@threads/core`	`exa()` from `threads.search`	Built-in tools
Tavily	`tavily()` from `@threads/core`	`tavily()` from `threads.search`	Built-in tools
Brave	`brave()` from `@threads/core`	`brave()` from `threads.search`	Built-in tools

Memory and knowledge

	TypeScript	Python	Guide
Local memory	`localMemory()` from `@threads/core`	`local_memory()` from `threads`	Memory
Supermemory	`supermemory()` from `@threads/supermemory`	`supermemory()` from `threads.supermemory`	Memory
Zep	`zep()` from `@threads/zep`	`zep()` from `threads.zep`	Memory
Local knowledge base	`localKnowledge()` from `@threads/core`	`local_knowledge()` from `threads`	Knowledge

There is a mem0() factory, but it refuses: the mem0ai SDK sends requests outside any transport threads can fence, so a call could not be recorded or replayed. Use supermemory() or zep(). See mem0.

Stores

The store is where the event log lives. Everything else is built on it.

	TypeScript	Python	Guide
SQLite	`sqlite()` from `@threads/core`	`sqlite()` from `threads`	Durability
Postgres	`postgres()` from `@threads/postgres`	`postgres()` from `threads.postgres`	Deploying

Telemetry

	TypeScript	Python	Guide
OpenTelemetry (OTLP)	`otel()` from `@threads/otel`	`otel()` from `threads.otel`	Observability

Writing your own

Every one of these is an ordinary implementation of a public protocol — Model, Sandbox, ChannelAdapter, MemoryProvider, KnowledgeProvider, SearchBackend, Store, Exporter — documented in the API reference. Nothing above uses a private hook, so a provider we don't ship is an adapter you can write yourself.

On this page