agent = ai.agent(tools=[get_weather])

messages = [ ai.system_message("You are a helpful weather assistant."), ai.user_message("What's the weather in Tokyo?"), ]

async for msg in agent.run(model, messages): print(msg.text_delta, end="")


`ai.openai(model_id)`, `ai.anthropic(model_id)`, `ai.ai_gateway(model_id)` — provider factories, callable, return `Model`. Clients auto-created from env vars (`AI_GATEWAY_API_KEY`, `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`). Pass `ai.Client(base_url=, api_key=)` to the provider call for custom endpoints: `ai.openai("gpt-5.4", client=c)`. `provider.list()` returns available model IDs.

`ai.stream(model, messages, ...)` — streaming without an agent loop. Returns `StreamResult` with `.text`, `.tool_calls`, `.output`, `.usage` after iteration.

`ai.generate(model, messages, params)` — non-streaming generation. `params` is `ImageParams` or `VideoParams`.

`ai.check_connection(model)` — verify credentials and model availability.

## Messages

Immutable Pydantic models. Use builders:

```python
ai.system_message("Be concise.")
ai.user_message("Describe this image:", ai.file_part(url))
ai.assistant_message(...)
ai.tool_message(...)       # merge one or more tool-result messages/parts
ai.tool_result(...)        # single ToolResultPart
ai.thinking(...)

Key properties on streamed messages:

msg.text_delta — current text chunk (for live display)
msg.text — full accumulated text
msg.tool_calls — list of ToolCallPart on assistant messages
msg.output — validated Pydantic instance (when using output_type)
msg.get_hook_part() — find a hook suspension part

Serialize: msg.model_dump(). Restore: ai.messages.Message.model_validate(data).

Streaming tools

A regular @ai.tool returns the tool result directly. An async-generator tool yields intermediate values that stream to the consumer as ai.events.PartialToolCallResult events; an aggregator decides what the model sees as the final result.

Declare the aggregator via the return-type annotation:

@ai.tool
async def fetch(url: str) -> ai.StreamingStatusTool[str]:
    yield "connecting..."
    yield "downloading..."
    yield body              # last yield = tool result; earlier yields = status

@ai.tool
async def render(prompt: str) -> ai.StreamingTextTool:
    async for chunk in stream:
        yield chunk         # all chunks concatenated into the tool result

@ai.tool
async def research(topic: str) -> ai.SubAgentTool:
    sub = ai.agent(tools=[...])
    async for event in sub.run(model, msgs):
        yield event         # final assistant text becomes the tool result

These aliases expand to Annotated[AsyncGenerator[T], ai.Aggregate(Aggregator)]. For a parameterized aggregator, write the marker directly:

type Joined = Annotated[AsyncGenerator[str], ai.Aggregate(ai.ConcatAggregator, delim="\n")]

Or pass the factory as a keyword argument: @ai.tool(aggregator=ai.LastAggregator). Specifying both an annotation marker and the kwarg raises TypeError.

Built-in aggregators: ai.ConcatAggregator, ai.LastAggregator, ai.MessageAggregator. Subclass ai.SimpleAggregator[Item, Result] (in/out same type) or ai.events.Aggregator[Item, Result, ModelResult] (separate model output) for custom ones.

Custom agent loops

Override the default loop when you need approval gates, routing, or batching.

The loop yields AgentEvents for the consumer and updates history with context.add(message). Anything not added to context.messages won't be visible on the next turn.

Concurrent form — fan tools out as the model emits them, mirrors the default loop.

my_agent = ai.agent(tools=[get_weather, get_population])

@my_agent.loop
async def custom(context: ai.Context):
    while context.keep_running():
        async with (
            ai.stream(
                context.model, context.messages, tools=context.tools
            ) as s,
            ai.ToolRunner(s) as tr,
        ):
            async for event in ai.util.merge(s, tr.events()):
                yield event
                if isinstance(event, ai.events.ToolEnd):
                    tr.schedule(context.resolve(event.tool_call))

            context.add(s.message)
            context.add(tr.get_tool_message())

Sequential form — drain the stream, then run tools. Simpler when you need to gate tool calls before executing them:

@my_agent.loop
async def custom(context: ai.Context):
    while True:
        s = ai.stream(
            context.model, context.messages, tools=context.tools
        )
        async for event in s:
            yield event
        context.add(s.message)

        tool_calls = context.resolve(s.tool_calls)
        if not tool_calls:
            return

        async with asyncio.TaskGroup() as tg:
            tasks = [tg.create_task(tc()) for tc in tool_calls]

        context.add(ai.tool_message(*(t.result() for t in tasks)))

Loop helpers: context.model, context.messages, context.tools, context.resolve(s.tool_calls), context.keep_running(), context.add(msg). await tc() executes a tool call and returns a ToolCallResult; pass those (or Messages) to ai.tool_message(...) to merge into one tool-result message.

Multi-agent

Use asyncio.gather with ai.yield_from(...) and labels to run agents in parallel:

async def multi(model: ai.Model, query: str) -> str:
    researcher = ai.agent(tools=[t1])
    analyst = ai.agent(tools=[t2])

    r1, r2 = await asyncio.gather(
        ai.yield_from(
            researcher.run(model, msgs1),
            label="researcher",
            aggregator=ai.MessageAggregator,
        ),
        ai.yield_from(
            analyst.run(model, msgs2),
            label="analyst",
            aggregator=ai.MessageAggregator,
        ),
    )
    return f"{r1}\n{r2}"

Each forwarded event is wrapped in ai.events.PartialToolCallResult carrying the label, so the consumer can route by source.

Hooks

Typed suspension points for human-in-the-loop:

class Approval(pydantic.BaseModel):
    granted: bool
    reason: str

Inside agent code (blocks until resolved):

approval = await ai.hook(
    "approve_send_email",
    payload=Approval,
    metadata={"tool": "send_email"},
)
if approval.granted:
    await tc()

From outside:

ai.resolve_hook("approve_send_email", {"granted": True, "reason": "User approved"})
await ai.cancel_hook("approve_send_email")

Hook messages have role="signal" with a HookPart.

Long-running mode (interrupt_loop=False, default): await blocks until resolved. Use for websocket/interactive UIs.

Serverless mode (interrupt_loop=True): unresolved hooks cancel the run. Pre-register a resolution with ai.resolve_hook(...) before rerunning.

Consuming hooks in the iterator:

async for msg in my_agent.run(model, messages):
    if msg.role == "signal" and (hook := msg.get_hook_part()):
        answer = input(f"Approve {hook.hook_id}? [y/n] ")
        ai.resolve_hook(
            hook.hook_id,
            Approval(granted=answer == "y", reason="operator"),
        )
        continue
    print(msg.text_delta, end="")

Structured output

class Forecast(pydantic.BaseModel):
    city: str
    temperature: float

stream = await ai.stream(model, messages, output_type=Forecast)
async for msg in stream:
    ...
stream.output.city

MCP

tools = await ai.mcp.get_http_tools(
    "https://mcp.example.com/mcp",
    headers={...},
    tool_prefix="docs",
)
tools = await ai.mcp.get_stdio_tools(
    "npx", "-y", "@anthropic/mcp-server-filesystem", "/tmp",
    tool_prefix="fs",
)

Returns Tool objects usable in ai.stream(...) or ai.agent(...).

AI SDK UI adapter

from ai.ai_sdk_ui import UI_MESSAGE_STREAM_HEADERS, to_messages, to_sse_stream

messages = to_messages(request.messages)
return StreamingResponse(
    to_sse_stream(agent.run(model, messages)),
    headers=UI_MESSAGE_STREAM_HEADERS,
)

Middleware

Subclass ai.Middleware and override the wrap methods you need. Pass to agent.run(..., middleware=[...]). Run-scoped, composable, first in list = outermost.

class LoggingMiddleware(ai.Middleware):
    async def wrap_model(self, call, next):
        print(f"calling {call.model.id}")
        result = await next(call)
        print("stream started")
        return result

    async def wrap_tool(self, call, next):
        print(f"tool {call.tool_name}({call.kwargs})")
        return await next(call)

async for msg in agent.run(model, messages, middleware=[LoggingMiddleware()]):
    ...

Five surfaces: wrap_agent_run, wrap_model, wrap_generate, wrap_tool, wrap_hook. Each receives a frozen context dataclass and a next callable. Use dataclasses.replace(call, ...) to modify before passing to next.

ai

ai

Quick reference

More from vercel-labs/py-ai

vercel-python-services

vercel-ai-sdk