The official Python SDK and CLI for the Yutori API — build agents that monitor, research, and browse the web, and operate computers with Yutori.
The SDK offers sync and async clients with full type annotations, plus a yutori CLI for authentication and managing resources from the terminal.
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.
On macOS or Linux, the recommended setup is the one-line installer:
curl -fsSL https://yutori.com/install.sh | bashInstalls 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.
Non-interactive install (CI, pipe, AI coding agent)
The SDK install, auth, and verification steps are skipped — auth needs a browser, verification needs an API key. MCP server and workflow skills install automatically without prompts.
To scope the MCP install to one coding agent, set YUTORI_INSTALL_CLIENT=<slug> (e.g. claude-code, codex, cursor). Unset, it registers for claude-code, codex, cursor, and gemini-cli. Run npx add-mcp list-agents for the full slug list.
Uninstall the CLI later
curl -fsSL https://yutori.com/uninstall.sh | bashRemoves the global yutori CLI. Saved credentials at ~/.yutori/ are left in place so they survive reinstalls — rm -rf ~/.yutori manually if you want a clean slate. Set YUTORI_UNINSTALL_ASSUME_YES=1 for scripted runs.
Install the package manually
pip install yutoriOr add it to an existing project with uv:
uv add yutoriAuthenticate manually
Run this once to save your API key:
yutori auth loginThis opens your browser to log in with your Yutori account and saves an API key to ~/.yutori/config.json. The SDK and CLI automatically pick it up.
If you installed the package with uv add, run uv run yutori auth login instead.
Or use an env var / pass the key explicitly:
from yutori import YutoriClient
client = YutoriClient() # Uses saved credentials or YUTORI_API_KEY
client = YutoriClient(api_key="yt-...") # Or pass explicitlyResolution order: explicit api_key > YUTORI_API_KEY env var > ~/.yutori/config.json.
Configure MCP server and skills manually
The installer sets these up automatically when Node.js is available. To do it manually:
npx add-mcp -n yutori "uvx yutori-mcp"
npx -y skills add yutori-ai/yutori-mcp -g -y -a claude-codeThe first command registers the Yutori MCP server with your editor. The second installs workflow skills scoped to your client — swap -a claude-code for another slug if needed (the skills CLI needs git on PATH).
The Yutori API provides four main capabilities:
| API | Description | SDK Namespace |
|---|---|---|
| Navigator | Browser- and computer-use models (Navigator n1.5, n2) | client.chat |
| Browsing | One-time browser automation tasks | client.browsing |
| Research | Deep web research using 100+ tools | client.research |
| Scouting | Continuous web monitoring on a schedule | client.scouts |
The Navigator API serves Yutori's models: Navigator n1.5 operates a webpage in a browser and Navigator n2 operates a complete desktop. Both take a task and a screenshot and return the next actions as tool_calls; your code executes them, sends the results back, and calls the model again until it indicates stop. The endpoint follows the OpenAI Chat Completions interface, so client.chat is a drop-in OpenAI-compatible client.
Capture a screenshot of the page and send it with the task:
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 run an agent loop: execute the returned actions on the page, capture a fresh screenshot, and call the model again until it returns text with no tool_calls. The complete agent loop lives in examples/navigator_n1_5.py, with custom-tool variants alongside it.
References and request options
The SDK defaults to Navigator n1.5 (n1.5-latest). Navigator n1.5 requests support selectable tool sets, disable_tools, and structured JSON output via json_schema (returned as response.parsed_json). See the Navigator reference for model IDs, parameters, and the full action space.
If you'd rather not manage browser infrastructure, use the Browsing API below, which runs the Navigator n1.5 on Yutori's cloud browser.
Navigator n2 operates a full desktop. It produces computer_batch calls — an ordered sequence of GUI actions — plus bash and file-tool (read/write/edit) calls. You implement the computer environment; the SDK's agent loop calls it to execute each action and sends the results back to the model:
from yutori import AsyncYutoriClient
from yutori.navigator import N2ComputerAgent
# Implement the async screenshot and input methods for your environment.
computer = MyComputer(...)
async with AsyncYutoriClient() as client:
agent = N2ComputerAgent(
computer=computer,
completions=client.chat.completions,
)
async for step in agent.run("Open Calculator and compute 17 * 23."):
... # each step yields the model's messages, tool calls, and tool resultsYour MyComputer environment is responsible for tool implementations (they are system-dependent); the N2ComputerAgent harness is responsible for tool transformation and batching:
| Tool | You implement | The SDK provides |
|---|---|---|
computer_batch |
the single-action GUI primitives (click, type, etc) |
the batching: validation, coordinate mapping, sequencing, the screenshot result. PointerKeyLifecycleMixin derives hold_key/wait/release_held_mouse_button from the primitives |
bash |
run_bash_command |
format_shell_output for the result format |
read/write/edit |
read_file/write_file/edit_file |
ShellFileToolsMixin: a complete implementation over a POSIX shell — inherit it and supply its two small hooks |
Three complete MyComputer implementations to copy from: direct_x11_adapter.py (direct X11), cua_adapter.py (API-backed sandbox), and MacOSComputer (native macOS). The first two inherit both mixins, so the only things they implement themselves are the GUI primitives and run_bash_command.
The full contract is documented in Navigator n2 loop.
The Cua cookbook instantiates this agent loop in a local Docker container:
yutori auth login # or export YUTORI_API_KEY=...
cd examples/navigator_n2
uv sync --python 3.12
uv run python local_docker.py --auto-approve "Open Calculator and compute 17 * 23"The script prints a Watch the desktop live: URL at startup — open it in a browser to follow along.
The Linux image in this example has no Calculator desktop app, and the example shows n2 moving between bash and GUI interactions to discover, launch, and visually operate the available xcalc fallback.
examples/navigator_n2_daytona.py runs the same agent on a Daytona Linux VM. Create an API key at app.daytona.io, then:
yutori auth login # or export YUTORI_API_KEY=...
export DAYTONA_API_KEY=... # https://app.daytona.io
uv run https://raw.githubusercontent.com/yutori-ai/yutori-sdk-python/main/examples/navigator_n2_daytona.py \
"Find the OS version and free disk space of this machine, and save a summary to a file on the desktop"See Run n2 on Daytona for a full walkthrough.
Drive your own local Mac
Yutori MCP ships the local harness, built on the same N2ComputerAgent:
uvx yutori-mcp computer-use setup
uvx yutori-mcp computer-use run "In Calculator, compute 17 * 23 and report the result." --app CalculatorReferences, model conventions, and long-run behavior
See the Navigator n2 reference for the tools, actions, and coordinate system, and the API reference for direct client.chat.completions.create(...) calls. A few conventions the model relies on — the bash tool rather than a GUI terminal, image-returning reads, the full tool set served together (tools can be disabled or added, but n2 performs best with all five) — are covered in the API reference, and the SDK ships the reference file-tool implementation (ShellFileToolsMixin) for any sandbox with a shell. Long runs are compacted automatically once the context grows, which is the regime the model performs best in for long-horizon work; pass compactor=None to disable, and the run instead continues past the 64K working context and stops cleanly at the 128k serving limit.
Agent-loop helpers
The yutori.navigator subpackage exposes optional helpers for typical agent loops:
| Helper | Purpose |
|---|---|
aplaywright_screenshot_to_data_url(page) |
Capture a Playwright screenshot as a Navigator-optimized WebP data URL. |
denormalize_coordinates(coords, width, height) |
Map the Navigator 1000×1000 coordinate space to viewport pixels. |
format_task_with_context(task, ...) |
Append location, timezone, and current date to a task message. |
format_stop_and_summarize(task) |
Ask the model to summarize when hitting max steps or an error. |
trimmed_messages_to_fit(messages, max_bytes, keep_recent) |
Drop older screenshots to stay under the API size limit. |
map_key_to_playwright(key) / map_keys_individual(keys) |
Convert Navigator n1.5 lowercase key names to Playwright format. |
yutori.navigator.tools |
Packaged JS reference implementations for Navigator n1.5 browser tool sets (extract_elements, find, set_element_value, execute_js). |
N2ComputerAgent / TOOL_SET_COMPUTER_USE_LATEST |
The stable Navigator n2 agent loop and current computer-use tool set. |
N2InlineCompactor / N2Compactor |
Default-on context compaction for long n2 trajectories (pass compactor=None to disable), and the protocol for a custom history rewrite policy. |
Full helper reference: api.md.
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.
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.
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"])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"])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())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}")# Authentication
yutori auth login # Log in via browser
yutori auth status # Show whether an API key is configured locally
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 usageRun yutori --help or yutori <command> --help for full options.
See examples/ for complete working examples: Navigator n1.5 browser loops, custom tools, and Navigator n2 on local Docker or Daytona infrastructure.
See CONTRIBUTING.md for development setup.
- 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
Apache 2.0 — see LICENSE.