Noema
    Noema
    Tooling

    Tool Usage

    Tools extend Noema beyond plain text generation. Arm the ones you need for the current chat, let the model call them when they improve accuracy, and review the returned output before relying on the result.

    Live information
    Web Search

    Best when the assistant needs current news, fresh references, or recent web content.

    Code execution
    Python Tool

    Best for calculations, statistics, parsing, algorithms, and other tasks where running code is more reliable than mental estimation.

    What tools do

    AI tools at a glance

    When a tool is armed, Noema can decide mid-conversation to call it. The result flows back into the model so the final answer can include live information, computed values, or structured execution output.

    Available tools

    What Noema currently exposes

    Web Search

    Current web information and source-backed answers.

    Network-enabled
    • Automatically crafts queries for recent news, facts, and references.
    • Summarizes multiple sources and cites them back in the reply.
    • Ideal when you need up-to-date information that is not stored locally.

    Python Tool

    Sandboxed Python execution for computational tasks.

    Local sandbox

    Use the Python tool when code execution will improve accuracy or save time, especially for calculations, statistics, parsing, data processing, algorithms, or other computational tasks.

    Enable & arm

    How users turn tools on

    Web Search

    1. Head to Settings → Tools and enable Web Search.
    2. Inside a conversation, open the tools menu and arm Web Search for that chat.
    3. Watch the transcript for tool start and finish notices.

    Python Tool

    1. Open Settings.
    2. Turn on Python Code Execution.
    3. In chat, arm Python from the + tool menu before sending a prompt.

    The setting is split into a global toggle in Settings and a per-chat armed state in the composer or tool menu.

    Python tool

    What the Python tool exposes

    Callable action

    The Python tool exposes a single callable action: noema.python.execute.

    It accepts one argument: code, which is Python 3 source code to run.

    The model should use print() for anything it wants returned in the output.

    Good for
    • Math and formula evaluation.
    • Statistics and numerical analysis.
    • Table and data transformations.
    • Parsing and text manipulation.
    • Algorithmic or STEM-heavy questions.
    • Any task where running code is more reliable than mental calculation.
    Sandbox & safety

    Python execution limits

    • Runs in a sandboxed environment with a 30-second timeout.
    • Has no network access.
    • Cannot access files outside a temporary working directory.
    • Blocks dangerous modules and system or process APIs.

    In practice, networking, subprocess spawning, and filesystem access outside the temporary sandbox are blocked. The tool is intended for controlled computation, not general system automation.

    Execution results

    What the model gets back

    Python execution returns structured output including stdout, stderr, exitCode, executionTimeMs, timedOut, and an optional error. That lets the model inspect the result and use it in the final answer.

    Availability rules

    When Python can be used

    • Python Code Execution is enabled in Settings.
    • Python is armed in the current chat.
    • A chat model is active.
    • The current model supports function calling.
    • No dataset is active or indexing.
    • The Python runtime is available on the device.
    Current limitations
    • Local MLX models do not expose the Python tool because tool calling is treated as unreliable there.
    • If a dataset is selected or currently indexing, Python is disabled.
    • If the current model does not support function calling, Python cannot be used.
    Recommended guidance

    When users should turn Python on

    Tell users to enable Python when they want the assistant to verify calculations or perform structured data work, and to leave it off for normal conversational prompts where code execution is unnecessary.

    Execution flow

    What happens under the hood

    1. Noema analyses your request and decides if a tool would help.
    2. It selects the best available armed tool and prepares the call payload.
    3. The tool executes and returns structured data.
    4. Results are woven into the final response with citations, summaries, or computed output.

    Privacy & control

    • Dataset queries stay on-device, and Python runs in a local sandbox.
    • Web Search is the tool that reaches the network for current information.
    • Disarm tools anytime from the chat header if you want to stay offline or avoid extra execution.