mistralai-vibe-sdk 0.2.0

Vibe SDK

Vibe SDK

High-level agent interface for the Vibe protocol.

Quick Start

from mistralai.vibe.protocol.patterns.completion.adapters.mistral import MistralCompletion
from mistralai.vibe.sdk import Agent, AgentConfig

completion = MistralCompletion.from_env(model="mistral-large-latest")
agent = Agent(
    config=AgentConfig(model="mistral-large-latest"),
    completion=completion,
)

# Async
async with agent.session() as session:
    async for state in session.run("Hello"):
        print(state)

# Sync
with agent.session_sync() as session:
    for state in session.run("Hello"):
        print(state)

Custom Local Tasks

To customize the underlying task implementation, pass an agent_task_factory. The factory receives the compiled AgentTaskConfig and the runtime completion, then returns any Task implementation. completion is still passed to Agent; the factory just receives it explicitly instead of capturing it through a partial.

from mistralai.vibe.protocol import AgentTask
from mistralai.vibe.protocol.patterns.completion.adapters.mistral import MistralCompletion

completion = MistralCompletion.from_env(model="mistral-large-latest")

def build_task(task_config, completion):
    return AgentTask(
        completion=completion,
        tasks={"my_tool": MyTask()},
        model_name=task_config.model,
        name=task_config.name,
        system_prompt=task_config.system_prompt,
        max_iterations=task_config.max_iterations,
    )


agent = Agent(
    config=AgentConfig(model="mistral-large-latest"),
    completion=completion,
    agent_task_factory=build_task,
)

Durable Tasks

Use the durable helper when you want Agent sessions to start workflow-backed tasks instead of local AgentTask instances. Install the workflow extra and import the helper from the workflow namespace.

from mistralai.client import Mistral
from mistralai.vibe.sdk.workflow import create_durable_agent_task_factory

client = Mistral(api_key="...").workflows

agent = Agent(
    config=AgentConfig(model="mistral-large-latest"),
    completion=completion,
    agent_task_factory=create_durable_agent_task_factory(
        client=client,
        task_queue="agents",
    ),
)