RunAnywhere

AI agent instructions

Read the Notte skill and follow its setup instructions when setting up Notte or building browser automation. This is the shared source for CLI setup, authentication, and the browser-to-code workflow.

RunAnywhere serves open models (DeepSeek, GLM, Kimi, MiniMax) through Wally Cloud, an OpenAI-compatible API. Point the open-source Notte framework at it to run act() and structured extraction in local sessions on those models, with no OpenRouter key. This applies to local sessions (notte.Session()). Cloud sessions use the platform’s server-side model configuration.

Prerequisites

Python 3.11+ and pip install "notte[llm]"
Chromium for local sessions: patchright install --with-deps chromium
A Wally Cloud key (sk-runa-...) from the RunAnywhere console

session.act() and session.check() are not in notte 1.10.1, the latest release; they ship in the next one. With 1.10.1, these settings apply to structured extraction.

Configure the endpoint

export NOTTE_LLM_BASE_URL=https://inference.runanywhere.ai/v1
export NOTTE_LLM_API_KEY=sk-runa-...
export NOTTE_LLM_MODEL=deepseek-v4.1-flash-fast

With no TYPESAFE_API_KEY or OPENROUTER_API_KEY set, act() uses this endpoint for both decisions and the text it types. If one of those keys is set, add NOTTE_ACT_DECIDER=llm to keep act() on RunAnywhere.

Example

runanywhere.py

from pydantic import BaseModel

import notte

# needs `pip install "notte[llm]"` and the NOTTE_LLM_* variables pointing at RunAnywhere


class Article(BaseModel):
    title: str
    born: str


with notte.Session() as session:
    _ = session.execute(type="goto", url="https://en.wikipedia.org/wiki/Main_Page")

    # act() decides each step and writes field values with the RunAnywhere model
    result = session.act("type Ada Lovelace in the search box")
    result = session.act("open the Ada Lovelace article")
    # a successful step is not proof of the destination: ask about the page itself
    if not result.success or session.check("Is this the Ada Lovelace article?") < 0.5:
        raise RuntimeError("the Ada Lovelace article did not open")

    # structured extraction uses the same endpoint
    article = session.scrape(instructions="Extract the article title and birth date", response_format=Article)
    print(article)

Choose a model

List the models your key can call:

curl https://inference.runanywhere.ai/v1/models -H "Authorization: Bearer $NOTTE_LLM_API_KEY"

Use the id of a model whose mode is chat. The example above was tested with deepseek-v4.1-flash-fast, deepseek-v4-flash-0731 and glm-5.3-flash. To use different models in act(), set NOTTE_ACT_MODEL for decisions and NOTTE_ACT_TEXT_MODEL for typed text. The text model defaults to the decision model, which defaults to NOTTE_LLM_MODEL. NOTTE_ACT_REASONING only applies to OpenRouter. On RunAnywhere, each model thinks at its own default level, which /v1/models reports under reasoning. Models that think by default take longer per step.

Integrations

Coding Agents

Agent Frameworks

Computer Use

Model Providers

Workflow Automation

Observability

Payments

File Storage

Email

Databases

Testing

Credentials

Browser Drivers

Infrastructure as Code

Browser Infrastructure

Prerequisites

Configure the endpoint

Example

Choose a model

​Prerequisites

​Configure the endpoint

​Example

​Choose a model

Prerequisites

Configure the endpoint

Example

Choose a model