Adapters Reference

Reference for the brain slot. For worked examples and guidance, see Bring Your Agent; this page is the contract.

The DialogAdapter protocol

Defined in unpod.adapters.base. Runtime-checkable - any object with these three methods passes isinstance(obj, DialogAdapter); no base class required.

class DialogAdapter(Protocol):
    async def turn(self, text: str, context: dict | None = None) -> str:
        """Return a complete response string. NOT called during live calls."""

    async def stream(
        self, text: str, context: dict | None = None, language: str | None = None
    ) -> AsyncIterator[str]:
        """Yield response tokens. HOT PATH - session.run() calls this on
        every user turn and streams tokens directly to the voice bridge.
        `language` is the per-turn language code (from `UserTextEvent.extra`),
        forwarded by `session.run()`. Custom adapters MUST accept the kwarg
        even if they ignore it."""

    def assist(self, text: str) -> None:
        """Inject a system instruction before the next turn."""

stream() now receives a language keyword argument on every turn. Any custom adapter whose stream() does not accept language=... raises TypeError on the first user turn. Add language: str | None = None to the signature.

session.run() calls only stream(). A single-chunk stream() fallback (await the full reply, yield once) produces choppy, high-latency audio with no error anywhere. See Streaming is the hot path.

Auto-wrapping

The ctx.session.dialog_machine setter accepts, in order:

A superdialog DialogMachine or LLMAgent - wrapped in a SuperDialogAdapter automatically.
Any object satisfying the DialogAdapter protocol - used as-is.
Anything else - raises TypeError.

Bundled adapters

All importable from unpod.adapters: AnthropicAdapter, DialogAdapter, HTTPAdapter, LangChainAdapter, MCPAdapter, OpenAIAdapter, SuperDialogAdapter.

Constructor signatures

Adapter	Constructor	Wraps
`OpenAIAdapter`	`(client, model="gpt-4o-mini", system_prompt=None)`	`openai.AsyncOpenAI` client
`AnthropicAdapter`	`(client, model="claude-haiku-4-5-20251001", system_prompt=None, max_tokens=1024)`	`anthropic.AsyncAnthropic` client
`LangChainAdapter`	`(chain, input_key="messages")`	LangChain Runnable with `ainvoke`/`astream`
`HTTPAdapter`	`(url, headers=None, timeout_s=10.0)`	A remote HTTP endpoint, any language
`MCPAdapter`	`(server_url, tools=None, llm="anthropic/claude-haiku-4-5", headers=None)`	An MCP server (preview - see below)
`SuperDialogAdapter`	`(dm)`	superdialog `DialogMachine` or `LLMAgent`

Streaming behavior

Adapter	`stream()` behavior
`OpenAIAdapter`	Real token streaming via `chat.completions.create(stream=True)`
`AnthropicAdapter`	Real token streaming via `messages.stream()`
`LangChainAdapter`	Real streaming via the chain’s `.astream()`
`SuperDialogAdapter`	Delegates to `dm.turn(text, stream=True[, language=…])` - `language` is passed through only when the wrapped machine’s `turn()` accepts it
`HTTPAdapter`	Single chunk - one POST round-trip, full reply yielded once
`MCPAdapter`	Falls back to `turn()`, which is not implemented yet

`assist()` semantics

Adapter	What `assist(text)` does
`OpenAIAdapter` / `LangChainAdapter`	Injects the instruction into the conversation history
`AnthropicAdapter`	Injects as a user/assistant message pair - the Anthropic API does not accept mid-conversation system messages
`HTTPAdapter`	Queues the instruction; sent as `system_instructions` in the next request body
`MCPAdapter`	Queues the instruction (pending full implementation)
`SuperDialogAdapter`	Delegates to the DialogMachine’s `assist()`

SuperDialogAdapter extras

Beyond the protocol, SuperDialogAdapter exposes the wrapped agent’s controls:

adapter.set_llm("anthropic/claude-haiku-4-5")        # swap the runtime model
adapter.switch_flow("billing", preserve_memory=False) # graph engine only (FlowSet)
adapter.is_complete                                   # reached a terminal state?
adapter.state                                         # current state (dict)

is_complete and state work on both engines; on the Playbook engine state returns {"checkpoint": ..., "slots": ..., "ended": ...}. switch_flow applies only to the legacy graph engine (a DialogMachine built from a FlowSet). See SuperDialog for the agent itself.

Error surfaces

Failure	What you see
Assigning a non-adapter to `dialog_machine`	`TypeError` from the setter, immediately
`HTTPAdapter` endpoint returns non-2xx	`httpx.HTTPStatusError` (the adapter calls `raise_for_status()`)
`HTTPAdapter` endpoint slower than `timeout_s`	`httpx.TimeoutException`
`MCPAdapter` without the `mcp` package	`ImportError` - install with `unpod[mcp]`
`MCPAdapter` with the package	`NotImplementedError` - full MCP orchestration is not implemented yet
`LangChainAdapter` with a chain expecting a different input shape	No error - the chain receives `{input_key: history}` and may silently misbehave; set `input_key` to match your chain
Lazy `stream()` in a custom adapter	No error - choppy, delayed audio on calls

Agent Studio

SuperDialog Framework

Embed in your stack

Speech runtime

Adapters Reference

The DialogAdapter protocol

Auto-wrapping

Bundled adapters

Constructor signatures

Streaming behavior

`assist()` semantics

SuperDialogAdapter extras

Error surfaces

​The DialogAdapter protocol

​Auto-wrapping

​Bundled adapters

​Constructor signatures

​Streaming behavior

​assist() semantics

​SuperDialogAdapter extras

​Error surfaces

The DialogAdapter protocol

Auto-wrapping

Bundled adapters

Constructor signatures

Streaming behavior

`assist()` semantics

SuperDialogAdapter extras

Error surfaces