Yutori Python SDK & CLI

The official Python SDK and CLI for the Yutori API — build web agents that autonomously execute tasks on the web.

The SDK offers sync and async clients with full type annotations, plus a yutori CLI for authentication and managing resources from the terminal.

AI agent install (recommended)

Paste this into Claude Code, Codex, Cursor, Windsurf, or another coding agent:

Use https://yutori.com/api/llms.txt and set up Yutori for me.

Manual install

On macOS or Linux, the recommended setup is the one-line installer:

curl -fsSL https://yutori.com/install.sh | bash

Installs the global yutori CLI via uv tool install and prompts to add the SDK to your project, run yutori auth login, register the MCP server, install workflow skills, and verify with a browsing task.

Python 3.9+ is required for the SDK.

curl -fsSL https://yutori.com/uninstall.sh | bash

pip install yutori

uv add yutori

yutori auth login

from yutori import YutoriClient

client = YutoriClient()                  # Uses saved credentials or YUTORI_API_KEY
client = YutoriClient(api_key="yt-...")  # Or pass explicitly

npx add-mcp "uvx yutori-mcp"
npx skills add yutori-ai/yutori-mcp -g

API Overview

The Yutori API provides four main capabilities:

Navigator API

The Navigator API hosts Yutori's visual-control models. Navigator n1 (n1-latest) and Navigator n1.5 (n1.5-latest) control browsers; the gated Navigator n2 preview (n2-preview) controls a complete desktop. Capture a screenshot, send it to the model, and execute the returned tool calls. The endpoint follows the OpenAI Chat Completions interface, so client.chat is a drop-in OpenAI-compatible client:

from yutori import AsyncYutoriClient
from yutori.navigator import aplaywright_screenshot_to_data_url
from playwright.async_api import async_playwright

async with AsyncYutoriClient() as client, async_playwright() as p:
    browser = await p.chromium.launch()
    page = await browser.new_page()
    await page.goto("https://www.yutori.com")

    image_url = await aplaywright_screenshot_to_data_url(page)

    response = await client.chat.completions.create(
        messages=[
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": "List the team member names."},
                    {"type": "image_url", "image_url": {"url": image_url}},
                ],
            }
        ],
    )

    message = response.choices[0].message
    print(message.content)  # Model's thoughts
    for tool_call in message.tool_calls or []:
        # Execute the requested browser action on `page`, append the tool
        # result to the conversation, capture a fresh screenshot, and call
        # the model again...
        ...

This snippet shows a single model call. In practice, you'll usually run an agent loop: execute the returned actions on the page, capture a fresh screenshot, and call the model again until it emits stop. Complete agent loops live in examples/.

The SDK defaults to Navigator n1.5 (n1.5-latest). Navigator n1 (n1-latest) is still supported for callers that want the older model. Navigator n1.5 adds selectable tool sets, disable_tools, and structured JSON output via json_schema (returned as response.parsed_json). See the Navigator n1.5 reference and Navigator n1 reference for model IDs, parameters, and the full action space.

Navigator n2 is a gated, non-streaming preview and does not change the SDK default. Pass an explicit dated tool_set; computer_use_tools-20260815 is the current macOS surface with nested batches, screenshots, local bash, and atomic click modifiers. n2 rejects caller-provided tools, disable_tools, json_schema, response_format, and non-auto tool_choice. The server preserves every screenshot in the two newest image-bearing messages (text-only messages do not consume a slot), strips older image parts while preserving other history, and inserts [Earlier screenshot omitted.] only when n2 pruning empties a message.

The Navigator API enforces a 10,000,000-byte complete-body limit. SDK trimming helpers use a separate 9,500,000-byte serialized-messages budget, reserving 500,000 bytes for the rest of the request. See the n2 Cua cookbooks for local macOS and disposable sandbox loops.

For direct macOS use, install the pinned driver extra and prepare the optional overlay before task startup:

python -m pip install 'yutori[macos]==0.9.0'
cua-driver permissions grant
python -c 'from yutori.navigator.macos import prepare_macos_overlay; prepare_macos_overlay()'

MacOSComputer then owns the persistent driver session, capture/input, shell lifecycle, cancellation, recovery, and presentation. Local shell execution remains disabled unless the caller passes allow_local_shell=True. The yutori-mcp computer-use setup command is the supported all-in-one installer and readiness flow.

Agent-loop helpers

The yutori.navigator subpackage exposes optional helpers for typical agent loops:

Full helper reference: api.md.

If you'd rather not manage browser infrastructure, use the Browsing API below, which runs the Navigator on Yutori's cloud browser.

Browsing API

Run one-time browser automation tasks on Yutori's cloud browser (or on Yutori Local with the user's logged-in desktop sessions):

task = client.browsing.create(
    task="Give me a list of all employees (names and titles) of Yutori.",
    start_url="https://yutori.com",
)

# Poll for completion
import time
while True:
    result = client.browsing.get(task["task_id"])
    if result["status"] in ("succeeded", "failed"):
        break
    time.sleep(5)

print(result)

Common options: require_auth=True for login flows, browser="local" for Yutori Local, webhook_url=... for async completion notifications. Failed tasks may include a rejection_reason.

client.browsing.list() enumerates your browsing tasks — omit limit to get them all, or pass status (running/succeeded/failed) and cursor to filter and paginate.

Structured output

Define the output structure with a JSON Schema dict or a Pydantic model:

from pydantic import BaseModel  # optional dependency

class Employee(BaseModel):
    name: str
    title: str

task = client.browsing.create(
    task="Give me a list of all employees (names and titles) of Yutori.",
    start_url="https://yutori.com",
    output_schema=Employee,  # Auto-converted to JSON Schema
    webhook_url="https://example.com/webhook",
)

The same output_schema pattern applies to client.research.create and client.scouts.create.

Research API

Perform deep web research using 100+ MCP tools (search engines, APIs, data sources):

task = client.research.create(
    query="What are the latest developments in quantum computing from the past week?",
    user_timezone="America/Los_Angeles",
)

# Poll for results
while True:
    result = client.research.get(task["task_id"])
    if result["status"] in ("succeeded", "failed"):
        break
    time.sleep(5)

Failed tasks may include a rejection_reason.

client.research.list() enumerates your research tasks — handy for exporting or recovering task IDs from a large batch. Omit limit to get them all, or pass status / cursor to filter and paginate:

completed = client.research.list(status="succeeded")
for t in completed["tasks"]:
    print(t["task_id"], t["created_at"])

Scouting API

Scouts run on a schedule to monitor the web and notify you when relevant updates occur:

scout = client.scouts.create(
    query="News, product updates, and announcements about Yutori AI",
    output_interval=86400,  # Daily (seconds, min 1800)
    webhook_url="https://example.com/webhook",
)

# Manage scouts
scouts = client.scouts.list(status="active")
client.scouts.update(scout["id"], status="paused")
client.scouts.update(scout["id"], status="active")
updates = client.scouts.get_updates(scout["id"], limit=20)
client.scouts.delete(scout["id"])

Async Usage

AsyncYutoriClient mirrors YutoriClient with async methods:

import asyncio
from yutori import AsyncYutoriClient

async def main():
    async with AsyncYutoriClient() as client:
        usage = await client.get_usage()
        scouts = await client.scouts.list()
        print(usage, scouts)

asyncio.run(main())

Error Handling

from yutori import YutoriClient, APIError, APIConnectionError, AuthenticationError

try:
    client.get_usage()
except AuthenticationError as e:
    print(f"Invalid API key: {e}")
except APIConnectionError as e:
    print(f"Connection failed: {e}")
except APIError as e:
    print(f"API error (status {e.status_code}): {e.message}")

CLI

# Authentication
yutori auth login      # Log in via browser
yutori auth status     # Show current auth status
yutori auth logout     # Remove saved credentials

# Scouts
yutori scouts list
yutori scouts create -q "monitor for news"
yutori scouts create -q "monitor for news" -i daily -tz America/New_York
yutori scouts get SCOUT_ID
yutori scouts delete SCOUT_ID

# Browsing
yutori browse list
yutori browse list --limit 20 --status succeeded
yutori browse run "extract all prices" https://example.com/products
yutori browse run "log in and continue" https://example.com/login --require-auth
yutori browse run "export dashboard data" https://example.com/dashboard --browser local
yutori browse get TASK_ID

# Research
yutori research list
yutori research list --limit 10 --status running
yutori research run "latest developments in quantum computing" -tz America/Los_Angeles
yutori research get TASK_ID

# Usage
yutori usage

Run yutori --help or yutori <command> --help for full options.

Examples

See examples/ for complete working examples, including Navigator n1/n1.5 browser loops and gated Navigator n2 desktop cookbooks.

Contributing

See CONTRIBUTING.md for development setup.

Documentation

• docs.yutori.com — API reference, model versions, and parameter details
• platform.yutori.com — usage monitoring, billing, and API keys
• api.md — SDK and CLI surface reference

License

Apache 2.0 — see LICENSE.